In late 2025, an open-source AI agent project called OpenClaw burst onto the scene, attracting millions of visitors and over 100,000 GitHub stars within a single week. ‘Your assistant, your machine, your rules’ – the slogan ignited enormous enthusiasm across the Global North’s technology communities. Almost simultaneously, several technology giants rushed to release their own ‘AI agent’ products, each claiming to revolutionise the very nature of knowledge work. For technology professionals in the Global North, this may represent yet another exciting cycle of technical iteration; but for researchers, movement leaders, and intellectuals in the Global South, a more fundamental question demands attention: when everyone is celebrating the technical breakthroughs of AI agents, very few are asking who owns and controls these tools, and whose agenda they serve.

This is not a rhetorical question. When a researcher in the Global South uses the ChatGPT chat interface to assist her writing, she is a consumer of AI – consuming a product designed by others, working under rules set by others, on platforms owned by others, with data selected by others. She can, in principle, paste her methodology into the prompt and save the output as a file – but these are crude workarounds, not systemic capabilities. The system’s behavioural logic remains opaque to her, the embedding of methodology is ad hoc and fragile, and knowledge accumulation depends entirely on manual effort outside the platform. Each conversation is an island; nothing compounds.

But if this researcher possesses a research system composed of dozens or even hundreds of agents – each with clearly defined responsibilities and quality standards, capable of automatically searching hundreds of web pages, processing hundreds of documents, analysing according to her chosen theoretical framework, and generating research reports with thousands of citations – then she is no longer a consumer but a builder. The system belongs to her, the methodology is embedded by her, the accumulated knowledge belongs to her institution, and the entire research process is transparent and under her control.

The leap from consumer to builder is a critical step for the Global South in reclaiming technological sovereignty.

The technical foundation upon which this leap depends is called the declarative multi-agent system. This article addresses three progressively deeper questions: what exactly is so-called ‘agentic AI’? How does one build such systems? And how does one combine them with one’s own work to determine what to build? In the previous article in this series, we proposed the AI4SS (artificial intelligence for social science) technical architecture – ‘one core, three repositories, four domains’. The declarative multi-agent system and its pattern language POMASA, the subject of this article, constitute precisely the ‘one core’ – the intelligent runtime – the engine that drives the entire AI4SS system. The system’s purpose is analytical: it searches, verifies, reasons, and synthesises. That its outputs take the form of text is a consequence of its method, not its objective. POMASA is not a multi-agent system itself – it is a construction guide for building one: a set of verified architectural patterns that tell you what to build, while the intelligent runtime determines how.

This is not theoretical speculation. The Global South Insights (GSI) project of Tricontinental: Institute for Social Research has already built and operates dozens of declarative multi-agent systems, the largest and most complex of which – the Research Production System – comprises several hundred declarative agents. What follows is drawn from our experience.

1. A New Form of Software – The Declarative Multi-Agent System

When people speak of ‘agentic AI’, the term is often shrouded in a fog of commercial hype. Once this fog is cleared, what emerges is a straightforward engineering reality: a multi-agent system (MAS) is, at its core, a form of software – not magic, not the embryonic stirring of general intelligence, but a new form of software with a clear conceptual framework, architectural principles, and engineering methodology. Grasping this point is the starting point for pulling AI back from hype into material reality.

1.1 From Long Prompts to Multi-Agent Systems: Why MAS Is Necessary?

Consider a researcher attempting to use AI for a complex research task. She might write a lengthy prompt, cramming role definitions, research methods, theoretical frameworks, output formats, and quality standards into a single block of text. As the task grows more complex, this prompt swells to 30,000 or even 50,000 words – and then three structural problems become impossible to avoid.

First, attention dilution. Even without exceeding the model’s context window, large language models confronted with lengthy, detail-laden prompts will simplify their execution autonomously, ignoring details they deem unimportant, rendering output quality unpredictable. Second, the absence of modular isolation. Even if the author mentally divides the prompt into ‘step one, step two’, all content is loaded into a single context, and modifying one section may unexpectedly affect another – much like writing all code in a single enormous function: though divided by comments into sections, it remains a monolith. Third, maintenance becomes intractable. Without flexible parameter passing, without dynamic process control, modifying a style guide requires reloading the entire 30,000-word prompt.

These are not hypothetical problems. Any researcher who has tried to encode a complete analytical framework – role definitions, source hierarchies, epistemological commitments, citation standards, style guides – into a single AI conversation has encountered exactly this: the AI begins ignoring citation standards; modifying the analytical framework destabilises the output format; each new conversation requires reloading everything from scratch. The carefully engineered knowledge specification has outgrown the single-conversation container.

For those with software engineering experience, this predicament is familiar. A long prompt is, in essence, text concatenation – role definitions, methodologies, style guides, and other components stitched together into one massive whole and fed to the AI in its entirety. It is akin to reading an entire operations manual aloud to someone from cover to cover, then expecting them to execute every task from memory. Software engineering was in this same stage in its earliest decades – all code written in a single file – and from the structured programming revolution of the late 1960s onwards, spent decades evolving toward today’s modular architecture: each module performs a distinct function, can be started and stopped independently, and collaborates through clearly defined interfaces. Current prompt engineering remains stuck at the ‘everything in one pot’ stage, and what multi-agent systems seek to accomplish is precisely to complete the journey that software engineering has already travelled – giving each module an independent lifecycle and clearly defined boundaries.

The solution: rather than entrusting everything to a single AI conversation, decompose a large task into multiple agents – each agent is an independent AI work unit, dedicated to a specific task (such as ‘collect literature on a given topic’ or ‘write analysis according to a specified framework’). Multiple agents, each responsible for its own role and working in coordination, constitute a multi-agent system.

1.2 Core Concepts of MAS

Understanding how multi-agent systems work requires grasping five core concepts.

The agent blueprint is a design-time definition document. It is a text file written in natural language – the most common format is Markdown (a lightweight formatting syntax widely used for writing documents), though even a plain text file will do – describing what an agent should do: role definition, input specification, task description, output specification, and quality standards. A blueprint is to an agent what an architectural drawing is to a building – the drawing is not the building itself, but a complete description of it.

The agent instance is a runtime execution entity. When a blueprint is ‘launched’, the system creates an instance with its own independent context window; it understands the task, executes operations, and produces results within its own space. A single blueprint can launch multiple instances simultaneously – just as a single set of blueprints can be used to construct many buildings of identical structure.

The runtime environment is the AI platform that executes agents. Unlike traditional programme runtimes, however, an AI runtime (such as Claude Code) is not a passive instruction executor but an intelligent runtime – it can understand natural-language intent, autonomously select execution strategies, and make judgements and adjustments when problems arise. A traditional runtime mechanically executes instruction sequences; an intelligent runtime understands intent and intelligently selects execution methods.

The data bus is the channel through which agents pass data to one another – in the most common current implementation, this is simply the filesystem. Agent A writes its output to a designated directory; Agent B reads from that directory as input. This seemingly ‘low-tech’ design embodies a profound consideration: all intermediate products are files that humans can directly read – complete, transparent, and traceable.

Reference data is externalised domain knowledge – methodological documents, theoretical frameworks, disciplinary standards, domain ontologies, and so on. For example, GSI maintains a reference document on the foundations of Marxist theory that is shared across multiple systems as a common analytical baseline (a future article in this series will examine this reference document in detail). Shared by all agents but invariant across tasks, reference data constitutes the knowledge foundation of a MAS and the vessel for institutional knowledge assets.

In practice, each of these five concepts maps onto concrete research activities. Agent blueprints are Markdown files describing tasks such as ‘collect and verify fiscal data on a country’s extractive sector’ or ‘analyse distributional impacts using a declared epistemological framework’. Agent instances are the individual AI work sessions launched from these blueprints – one instance collecting data on natural gas contracts, another simultaneously collecting data on mineral royalties. The runtime environment is the intelligent AI platform executing these agents. The data bus is a set of folders – one containing verified data, another containing thematic analyses, a third containing draft sections – each readable and inspectable at any moment. And reference data comprises the institutional knowledge assets assembled by the research team: epistemological framework documents, methodology guides, style standards, and the primary sources that ground every agent’s work.

1.3 Declarative versus Imperative: Why the Declarative Approach Is More Advanced?

There are two fundamentally different approaches to constructing multi-agent systems.

The imperative approach was the dominant method in pre-2025 frameworks (such as LangChain and AutoGen): agent behaviour is defined through Python code – code that controls how the agent acts. This approach requires programmers, imposes a high barrier to entry, demands substantial code, and makes debugging complex.

The declarative approach is a new paradigm: agent behaviour is defined through natural-language documents – describing what properties the result should possess, while delegating the how to the intelligent runtime to determine autonomously at execution time. After Anthropic released Claude Code in 2025, the declarative method rapidly became the prevailing approach – because Claude Code, as an intelligent runtime, natively supports launching and managing sub-agents through natural-language blueprints, rendering imperative frameworks no longer a prerequisite for building multi-agent systems.

This distinction is not a matter of stylistic preference but of architectural choice, and its essence lies in deferring decisions from design time to execution time. An imperative system determines at the moment the developer writes the code how to handle a given situation; a declarative system describes only what the result should look like, delegating specific execution strategies to the intelligent runtime to determine autonomously when confronting actual conditions. This is not laziness but an acknowledgement of a material fact – at design time, the developer possesses far less information than the AI system does at runtime. Just as an architect should not mark on the blueprints ‘if it rains tomorrow, use Plan B’, but should delegate tactical adjustments to the intelligent construction team on site.

The practical difference is stark. Under the imperative approach, a researcher would need a programmer to write Python code specifying, step by step, how to search for data, how to parse the results, how to handle a broken link or a PDF that fails to download. Under the declarative approach, the researcher writes a blueprint in plain English stating: ‘Collect fiscal data on the extractive sector from 2015 to 2025, prioritising government budget documents, reconciliation reports, and IMF Article IV consultations; verify each figure against at least two independent sources; flag any discrepancy exceeding 5 per cent.’ She describes what the output should look like; the intelligent runtime determines how to achieve it. The barrier between researcher and working system is no longer code – it is the clarity of her own specification.

Quantitative evidence from practice supports this choice. PayPal’s large-scale practice in 2025 found that the declarative approach reduced development time by 60%, tripled deployment speed, and decreased agent blueprint size by an order of magnitude. PayPal’s engineers concluded: ‘Most agentic workflows are composed of common patterns – data serialisation, filtering, retrieval-augmented generation, API orchestration – that can be expressed through a unified declarative language rather than requiring imperative code.’

History provides the best framework for understanding this transformation. From assembly language to high-level languages, from procedural programming to SQL’s declarative queries, from manual deployment to ‘infrastructure as code’ – each elevation in the level of abstraction has dramatically expanded the boundary of who can participate. The declarative multi-agent system is the latest milestone on this evolutionary line: it lowers the threshold for ‘building AI systems’ from ‘being able to write code’ to ‘being able to describe a problem’.

This means: the builder no longer needs to be a programmer.

What the builder needs is domain knowledge – the capacity to understand problems and to articulate clearly ‘what I need’. For researchers and movement leaders in the Global South, this transformation carries a decisive implication: the dependence on software engineers to build one’s own AI tools is no longer necessary.

2. Anyone Can Build – Pattern Language as a Construction Guide

The declarative multi-agent system lowers the threshold of who can build, but how to build well still requires best practices. A builder without experience, confronted with a blank page, still does not know where to begin, how many agents to define, how they should collaborate, or how quality should be assured. Pattern language exists precisely to solve this problem.

2.1 From ‘Being Able to Code’ to ‘Being Able to Describe’ – A New Empowerment

The core liberation of the declarative multi-agent system lies in this: what the builder needs to do is not write code but write documents. An agent blueprint is a Markdown file containing role definitions, input specifications, task descriptions, output specifications, and quality standards. Modifying an agent’s behaviour means modifying a passage of text – no compilation, no deployment, no debugging.

This means domain experts can directly participate in defining and modifying the system. A political economist can read an agent blueprint, understand what it does, and modify its analytical framework – in the natural language she knows well, rather than in Python, which she does not. Research methodology can be translated directly into system behaviour.

This carries particular significance for the Global South. Under the traditional AI development model, researchers must ‘translate’ their requirements for software engineers, who then implement them in code. In this translation process, intent is inevitably distorted and simplified. The declarative system eliminates this translation step – the researcher herself is the builder of the system.

2.2 Pattern Language: Executable Best Practices

But freedom does not automatically produce excellence. New builders still need to know ‘what a good system looks like’.

Pattern language is the classic answer to this question. It originates in architecture – Christopher Alexander proposed in his seminal 1970s work A Pattern Language a set of practice-verified design patterns, each solving a recurring design problem, with patterns forming an organic network of relationships that together guide builders in creating high-quality works.

Consider one of Alexander’s original examples: Pattern #159, ‘Light on Two Sides of Every Room’. Alexander observed that when people can choose, they always gravitate toward rooms with windows on two sides, while rooms with only one window are neglected – because single-sided light creates sharp contrasts and glare that instinctively cause discomfort. The pattern’s guidance is therefore: when arranging each room, ensure it has exterior walls on at least two sides where windows can be placed, allowing natural light to enter from more than one direction. This is a ‘pattern’ – it identifies a recurring problem (single-sided lighting causes discomfort), provides a verified solution principle (light on two sides), and does not prescribe specific construction methods. Software engineering successfully imported this thinking in the 1990s – the Gang of Four design patterns remain required study for every programmer.

POMASA is a pattern language that inherits this tradition from Alexander. Consider one of POMASA’s examples: Pattern STR-02, ‘Filesystem Data Bus’. When multiple agents collaborate, a central question is: how do they pass information to one another? One could use a database, a message queue, or various sophisticated technical solutions – but intermediate results then become invisible to humans, locked within the system’s internals. This pattern provides the principle: use the simplest possible mechanism – files and folders. Agent A writes its output as a file in a designated directory; Agent B reads from that directory as input. All intermediate products are documents that a human can open and read directly. If anything goes wrong at any stage, opening the corresponding folder reveals exactly what each agent produced – no black box.

In the AI era, pattern language undergoes a critical transformation. Traditional pattern languages were written for humans to read – humans understand the pattern and then translate it into code themselves. But in an intelligent runtime environment, the pattern language itself is written in natural language, and the AI runtime can understand natural language – therefore, the pattern language itself becomes executable. One can hand a set of patterns to an AI system, and it can directly generate a working multi-agent system from them.

This redefines what ‘open source’ means. Traditional open source meant publishing code for others to run.

In the AI era, publishing a pattern language is equivalent to publishing source code – because others can hand the patterns to an AI system and reconstruct an equivalent architecture.

2.3 POMASA: Our Pattern Language

POMASA (Pattern-Oriented Multi-Agent System Architecture) is precisely such a pattern language. It is not a specific software system but a set of architectural patterns refined from practice and verified through deployment.

POMASA comprises twenty patterns in four categories:

• Core patterns (COR, 2 patterns): Prompt-Defined Agent (COR-01) and Intelligent Runtime (COR-02) – the foundation of everything.
• Structure patterns (STR, 9 patterns): Reference Data Configuration, Filesystem Data Bus, Workspace Isolation, Business-Driven Agent Design, Composable Document Assembly, Methodological Guidance, and others – defining the system’s skeleton.
• Behaviour patterns (BHV, 6 patterns): Orchestrated Agent Pipeline, Faithful Blueprint Instantiation, Parallel Instance Execution, Progressive Data Refinement, and others – defining how the system operates.
• Quality patterns (QUA, 3 patterns): Embedded Quality Standards, Embedded Quality Standards, Layered Quality Assurance, Verifiable Data Lineage – defining how the system ensures reliability.

POMASA architecture overview: from blueprint definition to intelligent runtime execution, connected through the filesystem data bus

Each pattern is classified by necessity into three levels: Must, Recommended, and Optional. The minimum viable configuration requires only six must-have patterns to get started; ten recommended patterns cover the majority of scenarios; additional optional patterns can be introduced progressively according to actual needs.

A new builder need not absorb all twenty patterns before beginning. The six must-have patterns are sufficient to transform a single overloaded conversation into a modular system: defining agents in Markdown (COR-01), running them on an intelligent runtime (COR-02), passing data between them through folders (STR-02), isolating each agent’s workspace (STR-03), embedding quality standards directly into each blueprint (QUA-01), and orchestrating agents in a pipeline (BHV-01). As confidence grows, recommended patterns can be introduced progressively – parallel execution to research multiple dimensions simultaneously, layered quality assurance to separate the agent that writes from the agent that reviews.

This pattern language was not designed in the abstract. It was refined from real production systems – starting from five patterns and gradually growing to twenty through repeated system construction and fault resolution. Every pattern corresponds to a mistake once made or a design challenge repeatedly encountered.

2.4 Start Now: Three Steps to Begin

POMASA is not an armchair exercise. If you wish to try it now, only three steps are required.

The prerequisite is an intelligent runtime that supports sub-agent invocation – such as Claude Code, Cursor, or Cline. After installing the runtime of your choice, run npx skills add eXtremeProgramming-cn/pomasa in the command line to install the POMASA skill (this command supports all the runtimes mentioned above). Then, describe in natural language what you wish to research – for example: ‘Help me create a multi-agent research system to analyse digital economy policies in Southeast Asia.’ The AI will guide you in refining your requirements, recommend an appropriate combination of patterns based on your description, and then automatically generate the complete set of system files.

No single line of code is required throughout the entire process. The generated system is a set of Markdown files – agent blueprints that you can read, understand, and modify. POMASA itself is open-sourced under the Apache-2.0 licence. Moreover, this is not a demonstration piece – it is the same pattern language that GSI uses in its daily work.

2.5 Evidence from GSI: The Research Production System in Practice

GSI’s Research Production System (RPS) is the most thorough validation of POMASA in practice. RPS comprises several hundred declarative agents and hundreds of quality gates, the overwhelming majority blocking – 100% declarative, without a single line of Python orchestration code. Every agent is a Markdown file.

One of RPS’s flagship products is the Country Conjuncture Report – a systematic study of the political-economic situation in different countries. Taking the report on an African country as an example, this report passed through a multi-stage pipeline – from literature acquisition and ingestion, conjuncture analysis, research planning, deep research (with multiple topics per batch and several agents working in parallel), comprehensive writing, quality review, publication preparation, to summary generation. The entire process handled over a hundred input documents across dozens of research topics and produced a comprehensive report containing thousands of citations. The system was completed in a matter of days. Completing a country conjuncture report of equivalent quality through traditional methods would require six to twelve months. The efficiency gain amounts to 100 to 200 times.

But efficiency is not the only measure. Quality assurance is the central question that any declarative MAS must answer – because AI systems ‘hallucinate’ (research has shown that over half of the academic citations generated by ChatGPT are fabricated, and even GPT-4 still exhibits an 18% fabrication rate). RPS addresses this through hundreds of quality gates and dozens of producer-validator pairs: the agent that generates content cannot review its own output; validation must be carried out independently by a separate agent starting from zero. The system also incorporates a multi-layer fact-checking process with character-level precision – in a completed research report, hundreds of endnotes achieved zero fabricated citations. In external evaluation, the system scored highly, with the quality assurance dimension receiving a perfect score and the system assessed as ‘a production-ready system with significant innovation’.

This is not theory. This is a production system already in operation, generating real research outputs.

3. A Complete Example – From One Sentence to a Research Report

The preceding two sections set out the principles of declarative MAS and the pattern system of POMASA, but the reader may well ask: what does all of this look like in practice? Let us walk through the complete process, from a single paragraph of requirements to a finished research report containing 85 citations.

A reader who had been following the US-China technology competition came across seven articles on the subject online, covering chip wars, rare earth contestation, renewable energy, and tariff policy. Each article addressed a different facet of the issue, but he wanted a comprehensive overview that brought these scattered analyses together, helping him understand the complex landscape systematically. His core requirement, entered into the POMASA user input template, occupied a single paragraph: ‘The US-China chip tech war – covering chips, AI, robotics, energy, rare earths, and related supply chain issues. Bring this all together and complement it with full supply chain analysis, including China’s manufacturing prowess and supply chain ascendancy in adjacent areas such as electric vehicles, batteries, solar panels, and precision tools.’ He provided the seven articles to the system as reference materials.

He fed this requirement into the POMASA generator. The generator guided him through several key parameters: data source preferences (priority to policy documents, peer-reviewed research, and credible industry reports, with attention to non-English sources particularly Chinese); output format (research overview, 5,000 to 10,000 words); and quality assurance level (the strictest tier available). The generator then automatically selected thirteen of the twenty available patterns and produced a complete multi-agent research system – seven agent blueprints, a full set of reference data, methodology documents, and a directory structure, all consisting of readable, modifiable text files.

The generated system operates as an eight-stage pipeline. The narrative architect reads all seven reference articles and designs the narrative structure along with a detailed research plan. Deep researchers are assigned by topic and execute in parallel – each topic launches an independent agent instance that collects materials from the internet. The data verifier checks every item of collected data individually: not a sample but 100 per cent verification, a mandatory requirement of POMASA’s quality patterns. Analysts distil arguments from the verified materials, organised by theme. Section writers draft each section according to the designated style guide. Finally, the quality auditor, operating as an independent role, conducts a comprehensive review from scratch.

From filling in the user input template to generating the complete multi-agent system took roughly half an hour. The system then ran – deep research, data verification, analysis, writing, and quality audit – for several hours, investigating hundreds of verified sources from the internet in the process. The final output was a report entitled ‘Breaking the Siege: China’s Technological Ascent and the Unravelling of US Tech Hegemony’. Exceeding 11,000 words and containing eighty-five inline citations with URLs, it covered semiconductor sanctions and indigenous breakthroughs, AI chip competition, rare earth countermeasures, renewable energy dominance, ‘dark factory’ automated manufacturing, and implications for the Global South. The quality audit scored highly, with argumentation and Global South perspective both receiving perfect marks. The auditor simultaneously produced a concrete revision checklist – identifying areas for refinement and citations requiring further cross-verification – while assessing the core analysis as ‘publication-ready’.

The pattern, not the particular instance, is what matters. The same process applies whether the domain is technology competition, extractive industries in East Africa, land reform in South Asia, or labour conditions in Central America. The requirements change; the domain knowledge changes; the system architecture remains the same: requirements in, system out, research produced – with every step traceable.

This was not a flawless output – the very purpose of a quality audit is to identify deficiencies. Final judgements, refinements, and structural revisions still require human involvement; this is the practical meaning of ‘human in the loop’. What warrants attention, however, is the process itself: one reader’s single paragraph of requirements and seven reference articles, combined with POMASA’s pattern system and without any programming, generated a complete research system that autonomously produced a substantial report with extensive traceable citations. This is precisely the leap from consumer to builder described at the outset of this article – not a metaphor, but a fact we have witnessed firsthand.

4. What to Build – Design Guided by the Nature of the Task

‘How to build’ is a technical question; ‘what to build’ is a strategic one. A common mistake is this: once one has the declarative multi-agent system as a hammer, every task begins to look like a nail. Tasks of fundamentally different natures require fundamentally different system designs. The blind pursuit of full automation is the most common and most dangerous trap.

4.1 The Cynefin Framework: Understanding the Nature of Tasks

Dave Snowden’s Cynefin framework provides a decision tool for judging ‘what kind of MAS suits this task’. It classifies tasks by the clarity of their causal relationships into four domains:

Cynefin framework

The Clear domain – causal relationships are transparent and the path is known. Examples include batch file downloads, format conversions, and data cleaning. These tasks are suited to fully automated pipelines requiring no human intervention.

The Complicated domain – causal relationships exist but require expert judgement, and analytical frameworks can be determined in advance. An example is conducting country assessments according to an established indicator system. These tasks are suited to parallelisable analytical pipelines: multiple agents simultaneously collect and analyse data along different dimensions, then consolidate results.

The Complex domain – causal relationships can only be identified in retrospect, and critical insights must come from humans. An example is conducting a political-economic analysis of the class structure of a given country – without understanding the historical context of racial dynamics, data analysis alone will miss the mark entirely. These tasks demand iterative human-machine collaboration: AI conducts exploratory research and material preparation, the human provides insights and theoretical frameworks, and AI then performs deep analysis under the human’s framework guidance.

The Chaotic domain – the required information does not exist in any accessible data source and must be created through human action. An example is fieldwork conducted in factories and villages to gather primary materials – information that does not exist on the internet, has no archival record, and is beyond the reach of AI. In these tasks, AI can serve only as an auxiliary tool for mechanical work such as transcription, note organisation, and background material retrieval.

Most real research projects span all four domains simultaneously, and recognising this is essential to designing an effective system. Consider a study of why extractive wealth has not produced broad-based development in a given country. Translating government budget documents and cleaning fiscal datasets are Clear-domain tasks – fully automatable. Collecting and analysing extractive revenue data along established dimensions (royalty rates, corporate tax receipts, production-sharing terms) is a Complicated-domain task – agents can work in parallel across commodities using a predetermined analytical framework. But the central analytical question – identifying the specific configuration of class forces, state-capital relations, and international financial pressures that produce this outcome – belongs to the Complex domain: it requires the researcher’s own political-economic judgement, something AI cannot supply. And the experiences of communities displaced by mining operations – the anger, the broken promises, the local political dynamics – belong to the Chaotic domain, accessible only through fieldwork that no AI system can perform. A well-designed system automates the first two domains, structures iterative human-machine collaboration for the third, and limits AI’s role in the fourth to transcription and background retrieval.

A simple ‘three-question decision tree’ can help make the determination: Does the information needed to complete this task already exist? (No – Chaotic domain.) Is the analytical path from information to conclusions clear? (No – Complex domain.) Does it require expert-level judgement? (Yes – Complicated domain; No – Clear domain.)

4.2 Domain Degradation: The Evolution of the Research Process

The deeper insight that the Cynefin framework reveals is this: the research process itself is a progression from the Chaotic/Complex toward the Complicated/Clear.

A research project typically begins in the Chaotic or Complex – one is uncertain what to study, which framework to employ, or where to begin. As understanding deepens, the injection of human frameworks degrades Complex tasks into Complicated ones: once analytical dimensions and assessment criteria are established, what remains is structured data collection and analysis, which AI can accomplish efficiently. As methods mature, Complicated tasks further degrade into Clear ones: assessment processes become codified and can be repeated in fully automated fashion.

A well-designed multi-agent system should support this evolution – progressively increasing the degree of automation as understanding deepens. This is also the most important design warning: do not force-automate every task. Forcing automation on Complex or Chaotic domain tasks typically produces large volumes of shallow, insight-free content. Recognising AI’s boundaries is as important as recognising its capabilities.

4.3 GSI’s Practice: Different Configurations for Different Tasks

GSI’s actual systems embody distinct configurations for each of these four domains.

Clear domain: the fully automated pipeline. A researcher has a batch of documents in a language she cannot read and needs them translated into one she can. This task once required professional translators; now it can be handed entirely to AI – the inputs are defined, the processing rules are standardised, the output format is clear. The characteristics of such tasks are: information locations are known, retrieval paths are clear, processing methods are standardised – they can be entrusted entirely to AI for automated completion.

Complicated domain: the parallelisable analytical pipeline. The Digital Sovereignty Index (DSI) assessment is the representative case in this domain. The assessment framework encompasses four dimensions and sixteen indicators, elaborated into thousands of assessment criteria. For each country, the system collects hundreds of pieces of raw evidence, which are filtered through deduplication and quality checks to yield high-quality items. The complete pipeline produces a structured assessment report in a matter of hours – assessments have been completed for all eleven BRICS countries. The characteristics of such tasks are: the analytical framework is already established (designed in advance by human experts), and what remains is large-scale structured data collection and analysis, which AI can accomplish efficiently in parallel.

Complex domain: iterative human-machine collaboration. RPS’s country conjuncture studies represent this domain. The system employs a universal nineteen-dimension analytical framework (spanning material foundations, class forces, and contradiction analysis), but when applying this framework to a specific country, the critical step is ‘contextualisation’ – identifying the principal contradictions specific to that country, something AI cannot accomplish automatically. Take South Africa as an example: without understanding the historical legacy of settler colonialism, without understanding how racial dynamics have shaped the particular form of class contradictions in the country, AI can collect as much data as it likes and still miss the essential point entirely. Such tasks demand human-machine collaboration – AI processes thousands of source materials and produces dozens of thematic research reports; the human, on this foundation, provides critical analytical judgements (such as determining that South Africa requires the addition of ‘the legacy of settler colonialism’ as a country-specific analytical dimension); AI then, under the human’s framework guidance, weaves the fragmented research into a coherent analytical narrative. The injection of the human framework is precisely the step that degrades a Complex task into a Complicated one.

Chaotic domain: AI as an auxiliary tool. Fieldwork conducted in factories and villages to gather primary materials – information that does not exist on the internet, has no archival record, and is beyond the reach of AI. In this domain, the human leads the entire process, and AI undertakes only mechanical auxiliary work such as audio transcription, note organisation, and background material retrieval. Yet even here, AI’s assistance has value – one doctoral researcher, after using AI to process interview recordings, ‘discovered in the transcripts information that he had not heard or noticed during the interviews themselves’.

5. Implications for the Global South – The Convergence of Declarative MAS and AI4SS

The preceding sections addressed the hammer itself – why we chose this hammer, how to forge it, what it produces in practice, and what kind of nails it drives. This section must answer a different question: why does this hammer matter especially for the Global South?

The answer lies not in the technology itself but in the structural alignment between the technology and the concrete conditions facing the Global South.

5.1 The Structural Obstacles Confronting the Global South

Before discussing AI’s potential, we must first confront the structural obstacles facing the Global South.

The class asymmetry of knowledge production. Scholars who serve the interests of capital possess abundant resources, time, and energy to produce knowledge – yet this knowledge serves not the interests of the people but functions as an instrument of exploitation and oppression. At the same time, cadres of popular movements lack both resources and time. They cannot dedicate extensive hours to reading, research, and writing. The production and consumption of knowledge is monopolised along class lines – a cruel irony.

The contradiction between information volume and cognitive bandwidth. Lenin demanded in The Tasks of the Youth Leagues (1920) that Marxists ‘enrich their minds with a knowledge of all the treasures which humankind has created’, and the Marxist method of totality requires an integrated understanding spanning economics, politics, culture, and ideology. But these treasures of knowledge are too vast in quantity, too rapid in change, and too specialised in their subdivisions, far exceeding the limits of individual cognition. The method of totality thus becomes an ideal that cannot be put into practice.

Language barriers and resource barriers. The greatest volume of knowledge is produced in English, and translation resources are acutely scarce. Academic resources and databases charge exorbitant fees, and progressive media lack both resources and personnel. For a researcher in the Global South to access the data and literature that a scholar in the Global North takes for granted often requires a disproportionate expenditure of time and money.

The North-South infrastructure asymmetry. AI systems, publishing institutions, and citation indexes are concentrated in the North. Researchers in the Global South are evaluated by Northern standards, publish in Northern venues, and have their work framed by Northern theoretical concerns. The asymmetry of digital infrastructure – from computing power to network connectivity to electrical supply – means that the AI wave risks widening rather than narrowing the North-South gap.

5.2 A Threefold Liberation

Drawing together the foregoing analysis, the declarative multi-agent system offers the Global South a threefold structural liberation.

From technology consumption to technology construction – epistemological sovereignty. Declarative MAS combined with a pattern language such as POMASA enables the Global South to build its own tools using its own domain knowledge, without waiting for the benevolence of Silicon Valley. In a declarative system, the blueprint is the source code and methodology is the programme – and what the Global South possesses is precisely its own methodology and domain knowledge. ‘Publishing a pattern language is equivalent to publishing source code’ signifies a new form of open source – we are not merely using the North’s open-source tools; we are creating the South’s own open-source assets.

From knowledge monopoly to the democratisation of knowledge – capability multiplication. A small team of a few people, combined with a declarative MAS, can produce outputs that previously required a large research institution – several hundred agents, thousands of citations in a single country report, efficiency gains of 100 to 200 times. This capability multiplication is of particular consequence for the Global South: it enables resource-constrained progressive institutions, trade union research departments, and small independent media to conduct rigorous, evidence-based research and content production. The capacity for knowledge production, once monopolised, is being returned.

From mechanical toil to intellectual creation – the liberation of the human being. AI takes over the mechanical labour of desk work – searching, organising, transcribing, preliminary analysis – liberating humans from the heavy burden of information processing. But the purpose of this liberation is not leisure; it is to allow humans to return to the field, to the grassroots, to the masses, to do what only humans can do: generate insights, point the way forward, and render final judgements. What Marx called the unity of theory and practice – praxis – should be completed as a closed loop within the same person. In our practice, we have come to recognise with ever greater clarity: ‘Human thought, human experience, human insight – these are the most essential, most luminous elements in the entire process of knowledge and content production.’ AI is not intended to replace these – quite the contrary, the entire value of AI lies in allowing the human mind to focus on those most essential tasks that AI cannot perform.

The arc described across this article – from the structural limitations of a single AI conversation, through the modular architecture of multi-agent systems, to the pattern language that guides their construction – traces a journey from consumption to construction. A researcher who begins by pasting her methodology into a chat window and ends by building a system of hundreds of coordinated agents has not merely adopted a new tool. She has reclaimed the capacity to shape her own means of knowledge production. Her epistemological framework becomes reference data; her methodology becomes an agent blueprint; her quality standards become embedded quality gates. She does not code; she describes. And what she describes is not a wish but a specification – grounded in her training, her domain knowledge, and her political commitment to the communities whose conditions she studies. The system she builds belongs to her institution; the methodology it encodes is her own; the knowledge it accumulates compounds from project to project.

5.3 Restitution: From ‘Users’ Back to Makers

The significance of declarative MAS extends beyond efficiency gains or capability multiplication – it touches upon a deeper historical process.

For the past four decades, the information technology industry has systematically transformed the vast majority of people from makers of tools into mere ‘users’. As Bonnie Nardi documented in A Small Matter of Programming, the graphical user interface severed the continuity between using and creating; software commodification closed off the pathways to building; developer communities evolved into knowledge guilds; ever-escalating complexity raised the barrier higher and higher. In the end, it was not merely skills that disappeared – the very concept of ‘I can build tools for myself’ vanished from people’s language and imagination. When something cannot be spoken, it cannot be thought. The dispossession of capability thus became permanent and self-sustaining.

The declarative multi-agent system reverses this undercurrent. In this architecture, blueprints are Markdown rather than code; domain knowledge is the programme; methodology is the software. Building an agent system capable of autonomous deep research no longer requires permission from the programmer’s guild – what it requires is the knowledge in the domain expert’s mind: along which dimensions to analyse a problem, from which sources to gather information, which epistemological framework to guide judgement. And this knowledge is precisely what the progressive forces of the Global South have never lacked.

This is not endowment; it is restitution.

The capacity for ordinary people to build their own information tools existed from the beginning – researchers in the Unix era wrote their own scripts; users in the personal computer era booted their machines to a BASIC command line. This capacity is not a novelty brought about by technological progress; it is an old right that was systematically taken away by four decades of industrial evolution. The declarative multi-agent system gives it back.

Interactions with large language models frequently begin with a measure of disappointment. A user poses a complex and potentially productive question, only to receive a response that is remarkably bland, hollow, or beside the point. This disparity is particularly acute in the domain of serious knowledge production: researchers who seek to employ artificial intelligence to analyse imperialist policies, document popular struggles, and disseminate counter-hegemonic narratives find that its outputs consistently fall short of expectations. Many conclude from this experience that artificial intelligence is unsuitable for rigorous intellectual work.

This conclusion, however widespread, rests on a misdiagnosis. The problem lies not in the tool, but in how it is used.

‘Prompt engineering’ has become the dominant framework for addressing this challenge, focusing on how to formulate language in order to elicit better responses from large language models. Consider the widely adopted CO-STAR framework — Context, Objective, Style, Tone, Audience, Response — in which four of six elements address matters of ‘form of expression’: how to say it, in what tone, to whom, and in what format. This is not without utility — form of expression matters, and we shall address it in due course. However, from the standpoint of serious knowledge production, this framework exhibits systematic blind spots: epistemological framework, information reserve, methodology, opinion and insight — dimensions that are decisive for the quality of any knowledge product — remain entirely absent from its purview.

To reduce the challenge of employing artificial intelligence to a set of techniques for ‘how to phrase things’ constitutes a fundamental misdiagnosis. A more accurate assessment is this: the root cause of most failed interactions with artificial intelligence lies not in the wording of the prompt, but in the user’s failure to systematically organise their own questions, contexts, positions, and purposes. In other words, this is not a problem of prompt engineering; it is a problem of knowledge engineering. AI in this framework functions as an analytical instrument that produces text as its output — not a writing assistant that generates analysis as a side effect. As Herbert Simon demonstrated with his concept of ‘bounded rationality’, human decision-making is constrained not by a lack of information but by the cognitive limits on processing it — a principle that applies with equal force to human-AI collaboration, where the quality of outcomes depends less on the model’s capabilities than on the researcher’s capacity to organise the problem.

Amara is a researcher at a progressive policy institute in Dar es Salaam, tasked with producing a report on why Tanzania’s natural gas and mining boom has not translated into broad-based development. She has a laptop, decent internet, access to an AI tool, six weeks, no research assistant, and a small budget. She does not code. When she first asks her AI tool to ‘write a report on Tanzania’s extractive sector’, she receives a generic overview indistinguishable from a consultancy brochure — accurate in its broadest outlines, useless for her purposes. The six questions that follow chart the path from that initial disappointment to a genuinely useful research product.

This article poses six critical questions that constitute a framework for the practice of knowledge engineering: What problem are we actually trying to solve? From what standpoint do we view the world? What does the AI need to know? How should the analysis proceed? What should the final product look like? And what is the researcher’s distinctive contribution? These six questions apply to any mainstream large language model, independent of any specific tool — they concern not technology, but how researchers themselves think about and organise the production of knowledge.

1. What Problem Are We Actually Trying to Solve?

Most disappointing interactions with artificial intelligence share a common characteristic: the user proceeds directly to specifics without establishing a strategic purpose. ‘Help me write an article about the US-China tech war’ — such a request is virtually guaranteed to produce a superficial survey, because the request itself is superficial. Artificial intelligence is not incapable of producing work of analytical depth; rather, it cannot determine in which direction the desired ‘depth’ should extend.

Problem orientation is not ‘providing context’; it is an exercise in intellectual leadership. It requires the researcher, before engaging in any dialogue with artificial intelligence, to answer a fundamental question: What am I actually trying to understand? This question appears simple, yet it is precisely the step that most users omit.

Consider a concrete research project. In a study of the US-China semiconductor technology war, the researchers did not vaguely request an ‘analysis of US-China tech competition’, but instead formulated five progressively deepening core questions: How is the United States employing technology denial regimes to maintain geopolitical hegemony, and with what effects? How is China responding to technological containment, and what does this reveal about its development model? What is the current state of China’s supply chain dominance in critical sectors, and what are its implications? How is the technology war reshaping global supply chains and trade relationships? What are the implications for the Global South?

Beneath each core question lay further sub-questions that demarcated the boundaries and depth of the analysis. Under the question concerning US technology denial, for instance, the research further inquired: What policies and mechanisms are being deployed? How do these measures compare with historical technology containment strategies? What are the immediate and longer-term consequences?

The same artificial intelligence tool, confronted with ‘write an article about the US-China tech war’ on the one hand, and with such a structured five-tier system of questions on the other, produces fundamentally different outputs. The former yields an encyclopaedic listing of information; the latter yields a political economy study with an analytical framework, with stratification, with a point of view. The difference lies not in the tool, but in the question.

Rather than asking the AI to ‘analyse Tanzania’s extractive sector’, Amara formulates core questions: Why has natural gas and mining revenue not translated into industrialisation or improved livelihoods? What fiscal and contractual arrangements govern extractive rents, and whose interests do they serve? How have successive resource nationalism policies — from Magufuli’s mining reforms onward — altered the relationship between foreign capital and domestic accumulation? What role do international financial institutions play in shaping extractive governance? Each question generates sub-questions; together, they transform a vague topic into a stratified research programme.

In more advanced research practice, problem orientation can become still more incisive. In another project in which the author participated, researchers were required, before commencing any investigation, to answer a fundamental question: What is the principal contradiction driving the current state of affairs? This was not a rhetorical gesture but a binding requirement — if you could not identify the principal contradiction, the research would not proceed. This appears to be a demanding standard, yet it compels the researcher, before any dialogue with artificial intelligence, to complete the most difficult and most essential intellectual labour: determining what the research truly seeks to reveal.

The quality of the question determines the ceiling of the answer. A precise, stratified, analytically embedded system of questions is itself a product of intellectual leadership. Artificial intelligence can help explore the various dimensions of a question, can survey relevant literature and data — but determining which questions are worth posing can only come from the researcher’s own theoretical formation, political commitment, and insight into the realities under investigation.

2. From What Standpoint Do We View the World?

In the employment of artificial intelligence for knowledge production, a dangerous misconception is proliferating: many users regard large language models as neutral analytical instruments, expecting them to deliver ‘objective’ assessments on any given subject. This supposed neutrality, however, is an illusion. The training data for large language models is drawn predominantly from the English-language internet — a domain long dominated by Western ideological perspectives. Research has demonstrated that even open-source models developed in China exhibit considerable Western ideological bias: geopolitically favouring Western positions, epistemologically marginalising scholars from the South, economically naturalising neoliberal prescriptions, and culturally centring Western historical narratives.

When researchers employ these tools without critical reflection, a subtle yet far-reaching process unfolds:

It is not you who are calibrating the AI, but the AI that is calibrating you.

If you request an analysis of a developing country’s economic predicament without specifying an analytical framework, the model will almost inevitably proceed from the assumptions of neoliberal economics — market efficiency, comparative advantage, governance reform — because these constitute the dominant analytical paradigm in its training data. You may believe you have received ‘objective analysis’; in reality, you have received the reproduction of the biases embedded in the training data.

The solution is not to attempt to render artificial intelligence ‘more neutral’ — this is neither possible nor desirable. Every analysis embodies a standpoint; the only question is whether that standpoint has been consciously chosen or unconsciously absorbed. Knowledge engineering requires that, before collaborating with artificial intelligence, you explicitly declare your epistemological framework.

In a research project in which the author participated, researchers were required, before beginning any analytical work, to declare in advance a set of theoretical framework documents as the epistemological foundation of the entire study. These documents included a Marxist political economy framework, an analytical framework based on the Eight Contradictions of the Imperialist ‘Rules-Based Order’, a theoretical summary of hyper-imperialism, and a framework for contemporary class analysis. These were not optional references but binding prerequisites: without a declared epistemological framework, the research would not commence.

When Amara asks the AI to analyse why extractive revenues have not produced broad-based development, the model defaults to the ‘resource curse’ literature — institutional weakness, Dutch disease, governance deficits. The analysis is not wrong in every particular, but it systematically obscures the structural dimensions that matter most: the role of multinational capital in shaping fiscal terms, the historically conditioned weakness of the Tanzanian state’s bargaining position, and the international financial architecture that channels rents outward. She corrects for this by declaring her framework explicitly — providing reference texts on dependency and unequal exchange, on the political economy of extractivism in Africa, and on Tanzanian economic policy since ujamaa. The AI’s subsequent analysis is not ‘more biased’; it is more honest about the standpoint from which it proceeds.

To ‘declare an epistemological framework’ does not mean simply instructing the artificial intelligence to ‘use Marxist analysis’ — such directives tend to produce hollow texts saturated with rhetorical platitudes. What proves genuinely valuable is the systematic articulation of the core principles, terminological system, and analytical priorities of the chosen framework. The same project drew a precise distinction regarding how to apply the epistemological framework ‘invisibly’:

Correct structural analysis proceeds as follows: ‘The concentration of advanced chip manufacturing in Taiwan creates leverage that shapes US policy options…’ — the epistemological framework guides the analysis towards power structures and material conditions, while the analysis itself unfolds through concrete facts and logical reasoning.

Incorrect rhetorical platitudes proceed as follows: ‘The imperialist contradictions of capitalism drive the US ruling class to…’ — this is not analysis but the substitution of theoretical terminology for analysis. The framework has been crudely imposed upon the surface rather than deployed as a lens for deeper investigation.

The following is an actual prompt fragment from this project specifying how to apply the epistemological framework ‘invisibly’:

### Apply Ideological Framework Invisibly

The ideological framework should:
- Inform analysis of power, interests, and structures
- Help identify what questions to ask
- Provide conceptual tools for understanding

The ideological framework should NOT:
- Appear as explicit rhetorical terms
- Replace evidence with assumption
- Predetermine conclusions
- Make analysis feel propagandistic

(The above fragment is drawn from the project’s analysis methods guide, which runs to approximately 960 words. The total volume of prompts and reference guides for this research project exceeds 27,000 words — these figures, along with the output statistics cited later in this article, are drawn from the internal documentation of research projects conducted by the GSI team. Those interested in obtaining the complete prompts may contact the GSI team.)

This distinction is of paramount importance: the epistemological framework should guide what kinds of questions you pose (Whose interests are being served? How do power relations operate? How do material conditions shape possibilities?), rather than predetermining the answers. The framework instructs the artificial intelligence from what angle to view the world, but the analysis itself must rest upon evidence and logic.

For researchers in the Global South, this question is especially urgent. When Western-centric narratives dominate the training data, and when researchers fail to proactively declare their own epistemological positions — an anti-imperialist analytical lens, a structural attention to relations of dependency, a foregrounding of Southern experience — then every passage that artificial intelligence produces may unconsciously reproduce the knowledge hegemony of the North. To leave one’s epistemological framework undeclared is to unconsciously accept the ideological standpoint embedded in the model.

It must be emphasised that this is not a matter of making artificial intelligence ‘biased’ — on the contrary, it is a matter of making it honest. Every analysis has a perspective; every study has a standpoint. Declaring an epistemological framework transforms an implicit, unconscious bias into an explicit, conscious choice of position. This constitutes the deepest expression of subjectivity in knowledge production: you must know from what standpoint you view the world before you can require the machine to assist your analysis from that standpoint.

3. What Does the AI Need to Know?

A prevalent misconception treats large language models as a kind of oracle — you pose a question, and it delivers an answer from an omniscient internal knowledge base. In practice, large language models are reasoning engines that operate upon the information provided to them. When input information is sparse, they can produce only generic, hollow responses; when input information is rich and specific, their reasoning capacities can be brought fully to bear. The quality and scale of the information reserve directly determine the depth of what artificial intelligence can produce.

The information reserve has two complementary dimensions: the researcher’s own accumulated knowledge, and new information obtained in real time through the internet.

Domain Knowledge: What You Bring to the AI?

Every researcher accumulates, through sustained practice, a substantial body of knowledge assets: theoretical frameworks, core concepts, literature surveys, historical background materials. When collaborating with artificial intelligence, these knowledge assets should not remain idle — they should be furnished to the artificial intelligence as reference documents, forming the foundation for its reasoning and analysis.

A critical principle obtains here: reference materials provided by the researcher must be preserved in their entirety, never summarised or compressed. You may consider a theoretical framework document of several dozen pages excessively long, and wish to have the artificial intelligence ‘summarise the key points’ before proceeding — this is precisely the wrong approach. You can never know in advance which detail will prove critical during analysis. The context windows of contemporary mainstream large language models are now sufficiently capacious to accommodate large volumes of text; the full exploitation of this capacity is a fundamental practice of knowledge engineering.

Equally important is the separation of instructions from reference materials. Your instructions — what you wish the artificial intelligence to do — should be concise and precise; your reference materials — what you wish it to ‘know’ — should be comprehensive and complete. Conflating the two either buries instructions beneath reference materials or compresses reference materials unnecessarily.

Information reserves at scale can yield impressive results. In a study on the 80th Anniversary of the Victory in the World Anti-Fascist War, researchers constructed an information reserve spanning ten thematic sections — from the economic foundations of the belligerent powers to the architecture of post-war betrayal — ultimately producing a 5,900-word analytical study containing 150 citations across military archives, economic histories, and diplomatic records in multiple languages. Output at this depth and density is possible only when supported by a rich information reserve.

Amara lacks a large institutional library, but she has knowledge assets from her training and prior work — and these prove decisive. She uploads the full text of Tanzania’s 2009 Mining Act and its 2017 amendments, the Natural Wealth and Resources Acts, relevant sections of the Five-Year Development Plan, the TEITI reconciliation reports, and two political economy analyses of East African resource governance her institute had previously produced. She resists the temptation to summarise first; she provides them whole. The difference is not incremental but categorical: with these documents, the analysis cites specific fiscal terms, identifies discrepancies between legislated and effective tax rates, and traces the gap between policy intention and developmental outcome.

Internet Search: Opportunities and Pitfalls

One of artificial intelligence’s major capabilities is its ability to search the internet in real time, accessing the most current information. This means that research need not be confined to existing knowledge reserves; it can draw upon the latest policy documents, statistical data, research reports, and news coverage as they appear. However, this capability is accompanied by a distinctive challenge: how do you determine whether the information that artificial intelligence retrieves is credible?

In the US-China semiconductor technology war research project, the researchers specified in advance a credibility hierarchy for data sources: official policy documents (high) → peer-reviewed research (high) → research institution reports (medium-high) → industry analysis (medium) → news reporting (requiring cross-verification) → personal blogs and anonymous social media (excluded). This hierarchy was not decorative — it directly determined the weight that artificial intelligence should assign to different sources during analysis. The project simultaneously required that critical data be corroborated by at least two independent sources.

The deeper problem is that of artificial intelligence ‘hallucination’ — the fabrication of data that appears plausible but does not in fact exist. This is not an occasional malfunction but a systematic weakness of large language models. The following is an actual prompt fragment from the same project, designed to guide the artificial intelligence in identifying its own hallucinations:

## Warning Signs of AI Hallucination

Be especially vigilant for:

1. Plausible-sounding URLs that don't exist
2. Fabricated statistics
   - Very specific numbers that can't be found
   - Statistics that "perfectly" support the argument
3. Invented organisation names
   - Similar to real organisations but slightly different
   - E.g., "China Automobile Association" versus
     "China Association of Automobile Manufacturers"

(The above fragment is drawn from the project’s data verification guide, which runs to approximately 1,300 words.)

Sound knowledge engineering practice demands that systematic scepticism be maintained towards all information returned by artificial intelligence — this is itself an epistemological stance.

Information Reserve as Ongoing Practice

It bears emphasis that the information reserve is not a one-time preparatory exercise but a continuous practice that runs throughout the entire research process. Documents curated in the course of one project — theoretical frameworks, data source guidelines, credibility assessment criteria — become reusable knowledge assets. As projects accumulate, information reserves grow richer, and the depth of support that artificial intelligence can provide increases correspondingly. This constitutes a positive feedback loop: the greater the investment in knowledge engineering, the greater the returns.

4. How Should the Analysis Proceed?

Many researchers, when employing artificial intelligence, confine themselves to specifying ‘what to do’ — ‘analyse this issue’, ‘write an article about that topic’ — without specifying ‘how to do it’. This is akin to engaging a highly capable research assistant while providing only a vague direction and no methodological guidance whatsoever. The result is predictable: artificial intelligence organises its analysis according to its own defaults, producing text that appears reasonable but lacks methodological self-awareness.

The fourth question of knowledge engineering requires the researcher to become a project manager: specifying not only the objective, but the path towards it.

Two complementary strategies present themselves. The first is to prescribe the process: to furnish the artificial intelligence with a clear, step-by-step workflow. In the US-China semiconductor technology war research project, the analytical method was explicitly prescribed in four stages:

The first stage was evidence mapping — identifying all verified sources, extracting key facts, and annotating each source’s perspective and credibility. The second stage was pattern identification — seeking recurring themes across sources, identifying areas of consensus and disagreement, and noting changes over time. The third stage was critical analysis — interrogating the relationship between stated justifications and actual outcomes, identifying whose interests are served, and examining ‘what is not being said’. The fourth stage was synthesis — developing coherent arguments from the evidence whilst acknowledging complexity and uncertainty.

More granularly, the project required each argument to follow a five-step construction: claim — evidence — reasoning — qualification — significance. The following is an actual prompt fragment specifying this five-step argument construction:

## Argument Construction

For each major argument:

1. Claim: State the argument clearly and specifically
2. Evidence: Present the supporting data with citations
3. Reasoning: Explain why the evidence supports the claim
4. Qualification: Note any limitations or counterarguments
5. Significance: Explain why this matters

### Quality Checklist for Arguments

- Is the claim specific and falsifiable?
- Is the evidence from verified sources?
- Is the reasoning logical and explicit?
- Are limitations acknowledged?
- Is significance explained?

(The above fragment is likewise drawn from the project’s analysis methods guide.)

No step may be omitted: you must state clearly what the argument is (claim), what supports it (evidence), why the evidence supports the argument (reasoning), under what conditions the argument holds or does not hold (qualification), and why the argument matters (significance).

This methodological prescription may appear mechanical, but it possesses profound epistemological value. The act of prescribing a process compels researchers to think systematically about their own analytical practice — you are forced to answer explicitly: Where should the analysis begin? Through what stages should it proceed? How is argumentative rigour to be ensured? Many researchers, at the moment they are compelled to articulate their methodology, examine for the first time the analytical habits they have long taken for granted.

Amara prescribes a four-stage workflow. First, map the fiscal architecture — legislated rates, effective rates, stabilisation clauses in mining development agreements — drawing on her uploaded documents and verified sources. Second, identify patterns across time: how have terms shifted between the liberalisation of the 1990s, Magufuli’s resource nationalism, and the present? Third, critically analyse the gap between revenue captured and developmental expenditure, interrogating whose interests the prevailing arrangements serve. Fourth, synthesise these findings into an argument about structural impediments to broad-based development. Without this sequence, the AI would have produced a thematic survey; with it, the analysis builds cumulatively, each stage drawing on the last.

The second strategy is task decomposition: breaking complex tasks into manageable sub-tasks. A sweeping research question — such as ‘analyse the restructuring of global semiconductor supply chains’ — if presented to artificial intelligence as a single request, will inevitably yield a superficial overview. If, however, it is decomposed into more specific sub-tasks — first analyse US export control policies, then analyse China’s indigenous substitution efforts, next examine the strategic positioning of third countries, and finally synthesise these analyses into an overall assessment — each sub-task receives more thorough treatment.

In more complex research practice, the methodology can extend to sixty or seventy steps. The ‘steelmanning and dialectical fortification’ phase alone, for instance, comprises over a dozen steps. Among them is a ‘triple hostile reader test’: requiring that one’s own argument be scrutinised from the perspectives of a capitalist critic, a state-socialist critic, and an anarchist critic, respectively — can you rebut the strongest objections? This degree of methodological refinement ensures that the analysis is not merely internally consistent within its own framework, but can withstand rigorous challenge from opposing standpoints.

An important inflection point is implicit here. From four steps to seventy, the growth in methodological complexity gives rise to a practical problem: a single prompt can no longer accommodate such extensive methodological prescriptions. When researchers must simultaneously manage dozens of analytical steps, each with its own input-output standards and quality requirements, the natural next step is to distribute the methodology across multiple specialised agents — this is precisely the problem that declarative multi-agent systems, which we have discussed in detail in a separate article, are designed to address. Yet even before entering the domain of multi-agent systems, the researcher who prescribes a clear methodology within a single dialogue can already achieve significant improvements in the analytical quality of AI-generated output.

5. What Should the Final Product Look Like?

The same body of knowledge, expressed in different forms and serving different purposes and audiences, becomes an entirely different thing. A piece of in-depth research analysis, written as an academic paper, takes one shape; rewritten as training materials for community organisers, it takes another; produced as a podcast script, it takes yet another. Traditionally, effective writing has required the researcher to master three elements simultaneously: a solid command of background knowledge, depth of analytical insight, and command of the appropriate register. All three are indispensable, which explains why capable writers have always been scarce — those who simultaneously excel in analysis and expression are not readily found.

Artificial intelligence alters this equation. If you have addressed the preceding four questions — problem orientation, epistemological framework, information reserve, and methodology — artificial intelligence can provide powerful assistance with form of expression. You need not be simultaneously an analytical expert and a master of prose. This does not mean, however, that you may neglect the specification of form — on the contrary, the more precisely you define your formal requirements, the more closely the AI’s output will approximate your expectations.

In the US-China semiconductor technology war research project, the output template prescribed the Tricontinental dossier style with sentence-level precision. The document type was specified as ‘Tricontinental-style research dossier, 5,000 to 10,000 words’. Sentence complexity was prescribed through positive and negative examples:

Appropriate sentence complexity proceeds as follows: ‘While the United States has framed its semiconductor export controls in terms of national security, citing the potential military applications of advanced chips, the broader economic and geopolitical motivations — maintaining technological supremacy and slowing a rising competitor — are evident in the scope and evolution of these restrictions.’

Excessive simplicity proceeds as follows: ‘The US says export controls are about security. But they’re really about competition.’

These two sentences convey approximately the same information, yet the former employs complex syntactic structure to express the nuanced relationship between surface justification and underlying motivation, while the latter reduces the complexities of geopolitics to a simple binary opposition. For an academic research report, only the former mode of expression can bear the precision demanded by the analysis.

Vocabulary was likewise subject to detailed specification: terms of analytical precision such as ‘hegemony’, ‘containment’, ‘industrial policy’, and ‘technological sovereignty’ were to be employed; rhetorical ideological terms were to be avoided — structural analysis was to speak for itself, rather than being replaced by labels.

Amara’s institute has asked for a policy report of approximately 8,000 words, aimed at Tanzanian parliamentarians and East African civil society organisations — an audience that demands analytical rigour but not academic jargon. She specifies accordingly: the register should be formal but accessible, employing terms such as ‘fiscal leakage’, ‘beneficiation’, and ‘local content requirements’ where analytically necessary, but avoiding specialised academic vocabulary that would alienate non-specialist readers. She provides a previous institute publication as a style reference and specifies that every factual claim must cite a source. The report should include a summary of recommendations — not because the AI will generate them (that is her work), but because the structural template must accommodate them.

In more complex research projects, researchers prepare multiple writing reference documents specifying, respectively, content structure, citation standards, technical writing conventions, and referencing formats. For example, it may be required not only that sources be cited, but that each endnote contain context and a verifiable claim — enabling readers to independently verify every factual assertion.

It should be noted that the examples above pertain exclusively to textual output. Were the product a video script, a podcast outline, an infographic, or another multimedia form, the corresponding format requirements would naturally differ entirely — yet the same level of detailed specification would be required. The critical point is not what is specified, but that specification is necessary. Vague formal requirements produce vague results, just as vague questions produce vague answers.

This question also yields a cumulative benefit. With a solid information reserve and methodology in place, the researcher can rapidly re-express the same body of knowledge in multiple forms. A researcher’s deep analysis of a given topic can be disseminated simultaneously through academic papers, policy briefs, training materials, and social media content. Artificial intelligence renders this ‘analyse once, express many times’ approach a practical reality — and this is of particular significance for resource-constrained research institutions in the Global South.

6.What Is the Researcher’s Distinctive Contribution?

The preceding five questions define the supporting structure of knowledge production — problem orientation prescribes the direction, the epistemological framework prescribes the perspective, the information reserve provides the raw materials, methodology prescribes the process of elaboration, and form of expression prescribes the shape of the final product. Yet if only these five questions were addressed, the entire system would remain inert. It requires an initiator to conceive it, a decision-maker to guide it, and a judge to assess the value of its outputs. This sixth question — what is the researcher’s distinctive contribution — is not a step in the knowledge production process but the force that drives and evaluates the entire process.

The cybernetics pioneer Norbert Weiner, in his 1950 work The Human Use of Human Beings, offered a profound observation: the first industrial revolution automated physical labour, yet it did not render human beings superfluous — it compelled them to concentrate on the intellectual labour that machines could not perform. Artificial intelligence is initiating an analogous transformation in the domain of intellectual labour. What it automates is routine intellectual work — information retrieval, material organisation, format adjustment, first-draft generation — and this is precisely what liberates the researcher’s time and energy, enabling a focus on the higher-order intellectual work that belongs uniquely to human beings: formulating original research questions, making complex strategic and ethical judgements, and synthesising knowledge across different domains into new insight.

In one project in which the author participated, an explicit institutional arrangement was established to this end. After a rigorous ‘steelmanning’ process — in which the artificial intelligence scrutinised arguments from multiple opposing perspectives — and a ‘triple hostile reader test’ — in which arguments were challenged, respectively, from the standpoints of a capitalist critic, a state-socialist critic, and an anarchist critic — the process was compelled to halt before implementing any revisions, awaiting the approval of the human researcher. Without explicit human authorisation, the research could not proceed. The rationale was straightforward: ‘To ensure that human beings maintain control over the direction of argumentation.’

The same project also explicitly enumerated those things that artificial intelligence cannot replace: theoretical innovation, original analysis, critical thinking, value judgement, and domain expertise. The outputs of artificial intelligence were defined as ‘not publication-ready, but an excellent starting point’. This positioning is essential — it acknowledges that artificial intelligence can enormously accelerate the process of knowledge production, while insisting that ultimate intellectual judgement remains with human beings.

By this stage, the AI has produced a well-structured draft — grounded in Amara’s documents, organised by her methodology, written in her specified register. Yet the draft lacks something only she can supply. She has spent two years attending parliamentary hearings on mining governance; she has interviewed community leaders in Mtwara who watched gas revenues bypass their region entirely; she knows which civil society organisations have credible data and which reproduce donor talking points. It is Amara who decides that the central argument must foreground the gap between legislated resource nationalism and the contractual reality of existing mining development agreements — a judgement requiring not only information but the political discernment to know what matters most. She rewrites the recommendations entirely, drawing on conversations no AI has access to, and overrules the AI’s cautious ‘on the other hand’ equivocations where the evidence plainly supports a stronger conclusion.

For researchers in the Global South, this sixth question carries particular political significance. The irreducibility of opinion and insight means that artificial intelligence cannot substitute for knowledge embedded in communities, cannot replace the witness borne to struggle, cannot replicate the understanding that comes from standing alongside the oppressed. For generations, the power over knowledge production has been concentrated in the well-resourced institutions of the Global North. The proliferation of AI tools creates possibilities for resource-constrained institutions in the South — but only on one condition: that researchers maintain a firm hold on decision-making authority. What questions are worth posing, whose side to take, which sources to trust, what methods to employ, to whom to speak — these are choices that artificial intelligence cannot and should not make on your behalf.

The ultimate purpose of mastering knowledge engineering is not to become absorbed in dialogue with machines, but to employ machines to liberate us for the most essential of human tasks — to go deeper into the field, to listen to the ‘silent voices’, to acquire the tacit knowledge and popular wisdom that cannot be digitised. Genuine knowledge production has always moved ‘from the masses, to the masses’ — artificial intelligence makes this path more solid to walk.

Conclusion: Six Questions, One Whole

The six questions advanced in this article do not constitute a checklist but an organic whole. Problem orientation shapes the epistemological framework — what you seek to investigate determines from what standpoint you must proceed. The epistemological framework determines which information is relevant — your analytical perspective dictates what data is worth collecting and which sources deserve trust. The information reserve is transformed into analysis through methodology — abundant raw materials require a precise process of elaboration to yield valuable insight. The outputs of methodology are conveyed to audiences through form of expression — the same analytical conclusions require different modes of expression to reach different readers. And opinion and insight — the researcher’s distinctive contribution — drives the entire process: it is what determines which questions are worth posing, which standpoints are worth maintaining, and whether the final product carries meaning.

Six weeks later, Amara’s report is complete. It is not the product the AI would have generated on its own — that initial output was a generic survey devoid of standpoint or analytical architecture. Nor is it the product she would have produced without the AI — working alone, she could not have mapped the full fiscal architecture, cross-referenced legislative provisions against contractual terms, and produced a polished 8,000-word report in the time available. The report is the product of a structured collaboration: Amara supplied the questions, the standpoint, the documents, the methodology, the formal requirements, and the final judgement; the AI supplied the capacity to process, organise, draft, and redraft at a speed no solo researcher could match. Knowledge engineering made this collaboration productive; without it, the same tool would have produced the same bland overview with which she began.

As researchers systematically apply these six questions, a practical transformation occurs naturally: prompts inevitably grow longer and more complex. When you must simultaneously convey problem orientation, epistemological framework, information reserve, methodological prescriptions, and formal requirements, a single prompt quickly becomes a guidance document running to thousands of words. When a single prompt can no longer accommodate sufficient depth across all six dimensions, the natural next step is to distribute different dimensions of the work across multiple specialised agents — each agent responsible for one dimension, coordinated by orchestration logic. This is precisely the problem that declarative multi-agent systems are designed to address, as we have discussed in detail in a separate article.

Yet the core message of this article concerns not the direction of technological evolution but a more fundamental recognition: knowledge engineering is not a technical skill but a practice of intellectual discipline. It demands that, before asking a machine to speak on your behalf, you first know clearly what you know, what you believe, and what you wish to say. The true value of these six questions lies not in their capacity to help you write better prompts — although they do achieve this — but in the fact that the process of answering them is itself the core of serious knowledge production.

Artificial intelligence has accelerated the process of knowledge production, but it has not altered its essential nature. The value of knowledge continues to derive from deep insight into reality, from the conscious choice of standpoint, from the commitment to standing with those under investigation. To master knowledge engineering is to better accomplish this work that belongs to human beings — not to surrender it to machines.

This study emerges from the work of Global South Insights (GSI), a research initiative within Tricontinental: Institute for Social Research dedicated to developing analytical frameworks and technological capabilities that serve the needs of peoples in the periphery and semi-periphery of the world system. Over the past two years, our team has grappled with a deceptively simple question: can AI be harnessed to strengthen – rather than supplant – the intellectual labour of social scientists working on the problems that matter most to the Global South?

The answer we have arrived at is neither naïve techno-optimism nor reflexive technophobia. We have developed what we call AI4SS (AI for Social Science), a comprehensive framework that, grounded in trustworthy research infrastructure, operates across three interconnected layers: an ontological layer that reconceptualises the fundamental categories of research in the age of AI; a methodological layer that translates these concepts into concrete technical architectures; and a practical layer that implements these architectures in functioning systems. This framework is not merely theoretical. It has been tested in the production of research on digital sovereignty, technology competition, and development policy – research that has already begun to circulate among scholars and policymakers across the Global South.

What distinguishes AI4SS from the proliferation of AI tools now marketed to academics is its explicit commitment to what we call epistemic sovereignty: the principle that researchers in the Global South should control the means of knowledge production, rather than remaining dependent upon infrastructures and frameworks developed in and for the Global North. This does not mean rejecting Northern technologies – such a stance would be both impractical and counterproductive. It means building the capacity to adapt, extend, and redirect these technologies in accordance with our own research agendas and intellectual traditions.

The pages that follow present the AI4SS framework in its entirety: the problems it addresses, the ideas it embodies, the architecture it deploys, and the applications it has already enabled. Our hope is that this study will serve not only as an introduction to our work but as an invitation to collaboration. The challenges facing social science research in the Global South cannot be solved by any single team or institution. They require collective effort, shared infrastructure, and solidarity across borders. We offer AI4SS as a contribution to that larger project.

1. The Crisis of Social Science Research in the Data Age

The contemporary social scientist confronts a paradox. Never before has so much information about human societies been available; never before have the tools for processing that information been so powerful. Yet the lived experience of researchers – particularly those working in under-resourced institutions across the Global South – is often one of frustration, fragmentation, and futility. The gap between what is theoretically possible and what is practically achievable grows wider with each passing year. To understand why this is so, we must examine the structural contradictions that define social science research in the current conjuncture.

1.1 The Bandwidth Paradox: Data Abundance versus Cognitive Limits

The first contradiction arises from the collision between the explosive growth of available data and the biological constraints of human cognition. Consider the situation facing a researcher studying, say, the political economy of extractive industries in West Africa. Relevant materials exist in multiple languages (English, French, Portuguese, Arabic, and numerous African languages); in multiple formats (government reports, corporate filings, news articles, academic papers, social media posts, satellite imagery, audio recordings of community meetings); and across multiple jurisdictions (national, regional, international). The total volume of potentially relevant information far exceeds what any individual, or even any team, could possibly read, let alone analyse.

Faced with this deluge, researchers have traditionally resorted to two complementary strategies: sampling and abstraction. Sampling means selecting a manageable subset of the available data according to some criteria (random, purposive, convenience sampling). Abstraction means reducing the complexity of the data through categorisation, coding, and quantification. Both strategies are legitimate and often necessary. But both also entail significant losses. Sampling means that potentially crucial evidence may be excluded; abstraction means that the granular texture of social reality – the specific words people use, the particular contexts in which events unfold – is smoothed away in favour of generic categories.

The result is a kind of reductionist imperative built into the very foundations of social science methodology. Not because researchers prefer simplification, but because the cognitive bandwidth of human beings – what Herbert Simon termed ‘bounded rationality’, our limited capacity to read, remember, and reason – forces us to simplify. We lose breadth in order to gain depth, or sacrifice depth in pursuit of breadth. The possibility of achieving both simultaneously – of conducting research that is at once comprehensive and granular – has seemed foreclosed by the hard limits of human information processing.

1.2 The Rigour Gap: Generative Fuzziness versus Computational Precision

The second contradiction emerges from the peculiar characteristics of the AI tools now most widely available. Large language models such as ChatGPT have captured the public imagination and rapidly penetrated academic practice. These systems excel at certain tasks: summarising texts, generating prose, and answering questions in natural language. But they suffer from well-documented limitations that render them problematic for serious scholarly work.

Chief among these is the phenomenon of hallucination: the tendency of language models to generate plausible-sounding but factually incorrect statements. Research has shown that over half of the academic citations generated by ChatGPT are fabricated, and even the more advanced GPT-4 still exhibits an 18% fabrication rate. A model asked to summarise a body of literature may invent citations that do not exist; asked to analyse a dataset, it may report statistical findings that bear no relation to the underlying numbers. These failures are not bugs to be fixed but features intrinsic to the probabilistic architecture of such systems. Language models are trained to predict the most likely next word in a sequence, not to ascertain truth or verify facts. Their outputs are fundamentally generative rather than analytical – they produce text that sounds right rather than text that is right.

This creates a profound disconnect between the capabilities of general-purpose AI tools and the requirements of rigorous social science. Academic research demands not only the generation of plausible narratives but their grounding in verifiable evidence; not only the identification of patterns but their validation through statistical inference; not only the articulation of arguments but their logical coherence and empirical adequacy. The language model that can write a serviceable first draft of a literature review cannot perform a regression analysis, verify that the sources it cites actually exist, or assess whether its conclusions follow from its premises.

We are thus confronted with a gap between what can be calculated and what can be understood – between the deterministic precision of statistical computation and the probabilistic fuzziness of language generation. Bridging this gap requires not merely better models but fundamentally different architectures: systems that can integrate the fluency of language models with the rigour of computational methods, the creativity of generative AI with the accountability of evidence-based reasoning.

1.3 The Disciplinary Archipelago: Methodological Isolation versus Interdisciplinary Imperatives

The third contradiction concerns the organisation of academic knowledge itself. The social sciences have long been divided into distinct disciplines – economics, sociology, political science, anthropology, history – each with its own theoretical traditions, methodological conventions, and institutional structures. This division of labour has been productive in many ways, enabling the development of specialised expertise and the accumulation of disciplinary knowledge. But it has also created what we might call a Methodological Archipelago: a landscape of isolated islands, each speaking its own language, each employing its own techniques, with limited communication between them.

Consider the challenge of assessing the development strategy of a Global South country. An economist might analyse the health of the economic structure through macroeconomic indicators – GDP growth rates, capital formation ratios, trade dependency. A political scientist might examine governance capacity, the formulation and implementation of industrial policy, and the constraints that international dependency relations impose on policy space. A sociologist might focus on structures of inequality, the stratification of labour markets, and the distribution of development outcomes across different social groups. A historian might trace how colonial legacies have shaped the present economic structure and institutional arrangements. Any single dimension yields only a partial picture; the complete judgement – whether a country’s development path is sustainable, whom it benefits, what structural forces constrain it – can only emerge through integration. But integration is precisely what disciplinary structures impede.

The problem is not merely intellectual but practical. Different disciplines use different software, data formats, and analytical protocols. The qualitative researcher working with interview transcripts in NVivo (qualitative data analysis software) cannot easily combine their analysis with the quantitative researcher’s Stata (statistical software) output. The historian reading archival documents has no standardised way to link their findings to the economist’s time-series data. Cross-disciplinary collaboration, when it occurs, typically requires laborious manual translation between methodological frameworks – a process so costly in time and effort that it happens far less often than it should.

What is needed is not the abolition of disciplinary specialisation but its transcendence through methodological interoperability: the capacity for tools and techniques developed in one field to be readily employed in another, for findings generated through one method to be systematically integrated with findings from other methods. Such interoperability cannot be achieved through appeals alone; it requires technical infrastructure that enables different methodological traditions to communicate.

1.4 The Labour Inversion: Mechanical Toil versus Intellectual Creation

The fourth contradiction is perhaps the most insidious because it concerns not the nature of research outputs but the allocation of research labour. Ask any working social scientist how they spend their time, and you will likely hear a litany of tasks that have little to do with the creative intellectual work that drew them to scholarship in the first place: formatting citations, cleaning datasets, searching for literature, converting files between formats, navigating bureaucratic systems. These tasks are necessary but not sufficient; they must be done, but doing them does not advance knowledge.

The result is a perverse inversion of value: researchers spend the bulk of their time on low-value mechanical labour while the high-value intellectual labour – theoretical construction, interpretive insight, the formulation of genuinely new ideas – is squeezed into whatever hours remain. This inversion is especially damaging in the Global South, where researchers often lack the institutional support (research assistants, administrative staff, technical infrastructure) that their Northern counterparts take for granted. A professor in Lagos or La Paz may spend more time on bureaucratic and technical tasks than a professor in London or Los Angeles, leaving less time for the research that could establish their reputation and advance their career.

The irony is that many of these mechanical tasks are precisely what machines ought to be good at. Computers excel at repetitive operations, pattern matching, format conversion, information retrieval. Humans excel at judgement, interpretation, creativity, and the kind of contextual understanding that comes from lived experience. As Norbert Wiener observed in The Human Use of Human Beings (1950), a rational division of labour would assign to machines what machines do best and reserve for humans what humans do best. Instead, we have the opposite: machines (which often hallucinate) are used for thinking, while humans perform the drudgery that machines could handle.

2. Towards a New Paradigm: The Ontological Foundation of AI4SS

The contradictions outlined above are not inevitable features of the human condition; they are historical products of specific technological and institutional configurations. As those configurations change, new possibilities emerge. The AI4SS framework represents an attempt to seize those possibilities – to reconceptualise the fundamental categories of social science research in light of what artificial intelligence now makes possible. This reconceptualisation operates at the ontological level: it concerns not merely the tools researchers use but the very nature of research itself – the relationships among researcher, data, method, and output.

We organise the ontological foundation of AI4SS as a three-tiered structure we call ‘One Foundation, Four Pillars, One Core’. The Foundation is Trustworthy Infrastructure, answering the question ‘why should we trust it?’ – it provides system-level quality assurance mechanisms that are the prerequisite for all capabilities to operate reliably. The Four Pillars are Full-Data Perception, Full-Dimensional Analysis, Full-Channel Output, and Full-Disciplinary Synergy, answering the question ‘what can it do?’ – four mutually independent yet mutually reinforcing functional capabilities, each responding to the structural contradictions identified above. The Core is Human-in-the-Loop, answering the question ‘whom does it serve?’ – it defines the relationship between AI and human researchers, and constitutes the value anchor of the entire system.

The ontological foundation of AI4SS: One Foundation, Four Pillars, One Core

2.1 The Foundation: Trustworthy Infrastructure

To apply artificial intelligence to serious academic research, we must first confront a fundamental problem: the hallucination phenomenon in general-purpose AI tools is not a software defect awaiting a fix but an architectural feature intrinsic to probabilistic language generation. The ‘rigour gap’ identified in the previous section will not disappear simply because models become larger or more capable. Addressing it requires not better models but system-level constraint architectures.

This is precisely what distinguishes AI4SS from the proliferation of general-purpose AI tools now flooding the market. We are not offering researchers a more intelligent chatbot; we are building a trustworthy research infrastructure – a system designed from the outset with academic standards as its binding constraints. Trustworthiness unfolds along several interconnected dimensions. Execution paths must be predictable and auditable: the system operates along predetermined tracks, AI is not permitted to skip steps or take shortcuts, and every intermediate step can be inspected and traced. Every assertion must be traceable to its original source: the evidential chain of academic research cannot be broken, and the path from final output to underlying data must be preserved in its entirety. Content production and quality verification must be separated: as in the peer review system of academic publishing, the agent that generates content cannot review its own output – independent verification agents must perform the checking. Critical conclusions require multi-source cross-validation: judgements that rely on a single source carry an inherent risk of systematic bias, and robust conclusions can only emerge from the convergence of multiple independent streams of evidence.

Without this foundation, the more powerful the four capabilities above it become, the greater the risk of producing unreliable content at scale. A system capable of perceiving vast quantities of data, conducting multi-dimensional analysis, and generating multi-channel outputs would, in the absence of quality assurance mechanisms, simply produce errors with greater efficiency and at greater scale. Capability and trustworthiness must be built in tandem – this is the fundamental premise of the entire AI4SS architecture.

2.2 Full-Data Perception: Expanding the Cognitive Horizon

The first pillar responds to the bandwidth paradox by fundamentally expanding the range of data that can be brought within the researcher’s purview. Full-Data Perception means not merely accessing more data but transforming the relationship between researcher and data such that comprehensiveness and granularity become simultaneously achievable.

This transformation operates along three axes. The first concerns data ownership and access. Traditional social science research has relied heavily on publicly available datasets – government statistics, survey repositories, published texts. But some of the most valuable information for understanding social phenomena exists behind various barriers: proprietary corporate data, confidential government records, paywalled academic literature, and the tacit knowledge embedded in field notes and interview recordings that researchers accumulate but rarely systematise. Full-Data Perception involves the construction of what we call Institutional Cognitive Moats: secure repositories of high-value data assets that become the foundation for sustained research programmes. Through appropriate protocols for access control, encryption, and local storage, institutions can build cumulative knowledge bases that compound in value over time.

The second axis concerns format integration. Social reality does not present itself in neatly structured spreadsheets. It manifests in texts and images, in audio recordings and video footage, in maps and diagrams and handwritten notes. Traditional research methods have struggled to integrate these diverse modalities, forcing researchers to analyse each format separately and synthesise the results manually. Advances in multimodal AI now make it possible to process text, image, audio, and video within unified analytical frameworks – to extract information from a photograph and relate it to information from a document, to transcribe an interview and link its contents to quantitative datasets.

The third axis concerns language. The majority of human knowledge is not in English, yet English dominates the infrastructure of global scholarship. For researchers in the Global South, this creates a double bind: either restrict investigation to English-language sources and miss crucial local materials, or invest enormous effort in translation and lose productivity on other fronts. Semantic alignment technologies based on vector embeddings now enable a different approach: the mapping of texts in different languages into a common semantic space where meaning can be compared and retrieved regardless of the original language. A researcher need not be fluent in Portuguese to search a Brazilian archive, nor in Arabic to analyse Egyptian policy documents. The language barrier does not disappear, but it becomes permeable.

Full-Data Perception, then, is not merely about having more data. It is about establishing a new relationship between the researcher and the totality of relevant information – a relationship in which the limitations of individual cognitive bandwidth are overcome through technological augmentation, while the granularity and contextual richness of original sources are preserved rather than abstracted away.

2.3 Full-Dimensional Analysis: Against Reductionism

The second pillar addresses the reductionist imperative that has historically constrained social science methodology. Full-Dimensional Analysis means conducting research that is comprehensive in scope without sacrificing depth of engagement – that analyses the full dataset rather than a sample, that preserves the micro-texture of social reality rather than smoothing it into aggregate categories.

This is not a rejection of sampling or abstraction as such. There are contexts in which sampling remains necessary (when the population is genuinely open-ended or inaccessible) and abstraction remains valuable (when the research question specifically concerns aggregate patterns). But these should be methodological choices, not technological necessities. The researcher should be able to choose between sampling and exhaustive analysis based on the demands of the inquiry, not the limitations of available tools.

Full-Dimensional Analysis involves several interconnected capabilities. Exhaustive Computation means processing the entirety of a dataset rather than a subset, identifying patterns that sampling might miss, detecting outliers and edge cases that would be excluded from statistical averages. Holographic Analysis means maintaining access to the original data even as higher-level patterns are identified – the ability to drill down from an aggregate finding to the specific instances that compose it, to move fluidly between the macro and micro levels of analysis. Iterative Feedback means establishing dynamic loops between empirical material and theoretical construction, allowing theory to guide data exploration while remaining open to revision in light of what the data reveal.

Perhaps most importantly, Full-Dimensional Analysis seeks to integrate what have traditionally been separate methodological traditions: the quantitative and the qualitative, the statistical and the interpretive. This integration is not a matter of simply combining numbers and narratives; it requires systems capable of mutual validation, where statistical findings can be checked against qualitative evidence and qualitative interpretations can be tested against quantitative patterns. The goal is not to dissolve the distinction between methods but to enable their productive interaction.

2.4 Full-Channel Output: Multiple Outputs from Unified Knowledge

The third pillar concerns the relationship between research and its audiences. Traditionally, the primary output of social science research has been the academic article or monograph – a format optimised for peer review and disciplinary credibility but often ill-suited to reaching wider audiences or influencing policy. Researchers who wish to communicate beyond the academy must engage in a separate process of translation, rewriting their findings in accessible language, reformatting their arguments for different media, adapting their presentations for different contexts. This translation work is time-consuming and often undervalued by academic institutions.

Full-Channel Output inverts this relationship between research and output. Rather than producing a single canonical text that must then be translated into other formats, the researcher works with a unified knowledge base from which multiple outputs can be generated as needed. The same underlying analysis can manifest as an academic paper for a peer-reviewed journal, a policy brief for government officials, a long-form article for an educated general audience, a presentation for a public lecture, even an audio or video production for broader dissemination. Each output is adapted to its intended audience and medium, but all draw from the same evidentiary foundation and maintain the same argumentative integrity.

This capability depends on several technical features. Adaptive Content Generation means the ability to adjust register, vocabulary, and rhetorical structure according to the target audience. Stylistic Emulation means the capacity to match specific conventions – the format of a particular journal, the house style of a particular publication, the presentational norms of a particular discipline. Multimodal Output means generating not only text but also visualisations, graphics, and potentially audio and video. And Traceable Provenance means maintaining what we call the ‘digital umbilical cord’ that links every output back to its evidentiary sources – the ability to verify any claim by tracing it to the underlying data.

Full-Channel Output thus serves both efficiency and accountability. It allows researchers to maximise the impact of their work by reaching diverse audiences through appropriate channels, while ensuring that this diversification does not compromise scholarly rigour or evidential grounding.

2.5 Full-Disciplinary Synergy: Methodological Interoperability

The fourth pillar addresses the methodological archipelago by establishing infrastructure for cross-disciplinary integration. Full-Disciplinary Synergy means not merely encouraging collaboration between disciplines but building the technical systems that make such collaboration practically feasible.

The key mechanism is Capability Encapsulation: the packaging of disciplinary methods as modular units that can be used by researchers from other fields without requiring mastery of the underlying technical details. An economist’s regression toolkit, a sociologist’s qualitative coding system, a historian’s archival search protocol – each can be wrapped in a standardised interface that exposes its functionality while hiding its implementation complexity. A researcher who needs to conduct a statistical analysis need not become a statistician; they need only know how to invoke the relevant capability and interpret its outputs.

This encapsulation unfolds at two levels. At the tool level, standard protocols such as the Model Context Protocol (MCP) enable heterogeneous systems to exchange data and commands in consistent formats – tools built by different teams, in different programming languages, for different purposes, can nonetheless work together within integrated workflows. At the knowledge level, the emerging skills ecosystem (such as the open agent skills platform Skills.sh) enables methodological knowledge to be packaged as one-click installable, cross-platform reusable capability units – an analytical workflow developed by one researcher can be directly invoked by a researcher on the other side of the world, much like installing a software plugin. MCP addresses how tools connect; the skills ecosystem addresses how knowledge circulates. Together, the researcher becomes not a specialist in any single method but an orchestrator – capable of assembling multi-method workflows appropriate to the question at hand.

Full-Disciplinary Synergy also involves Fractal Collaboration: the ability to scale cooperative research from individual projects to large multi-institutional programmes. The same infrastructure that enables a single researcher to invoke multiple methodological tools enables teams to coordinate their work, institutions to share their capabilities, and networks of researchers across the Global South to pool their resources. The vision is one of cumulative, collective capacity-building rather than fragmented, duplicative effort.

2.6 The Core: Human-in-the-Loop – Augmentation, Not Replacement

As the core of the entire framework, Human-in-the-Loop is in some ways the most fundamental idea, for it concerns the relationship between human researchers and artificial intelligence itself. Human-in-the-Loop is both a technical architecture and a philosophical commitment: the principle that AI systems should augment human intelligence rather than substitute for it, should serve human purposes rather than impose their own.

This principle has several dimensions. Value Anchoring means that human researchers retain authority over the ethical frameworks, aesthetic standards, and social concerns that guide inquiry. The AI system does not determine what questions are worth asking or what answers are desirable; it operates within parameters set by humans according to human values. Cognitive Sovereignty means that humans retain ultimate veto power over the outputs of AI systems, particularly in domains requiring theoretical construction, meaning interpretation, and normative judgement. Every output generated by AI is a proposal subject to human review, not a final product; the system must make its reasoning processes visible and its evidential chains accessible, enabling researchers to verify and, where necessary, overturn algorithmic conclusions. The machine may propose; the human disposes.

At the same time, Human-in-the-Loop mandates the automation of mechanical labour – the delegation to machines of precisely those tasks that machines can perform reliably and that consume human time without engaging distinctively human capacities. Literature search, citation formatting, data cleaning, file conversion: these are activities that AI systems can handle more efficiently than humans and whose automation frees researchers to focus on work that genuinely requires human intelligence. But here a critical distinction must be drawn: the goal of automation is not an end-to-end unmanned pipeline but an iterative cycle of human-machine collaboration. AI4SS favours a Plan-Do-Check-Adjust iterative process in which human researchers maintain the capacity to intervene at every stage – to reassess direction, modify parameters, and pursue unexpected leads that the system alone would not have identified.

The ultimate goal is what we call Re-Embedding Experience: liberating the cognitive bandwidth of researchers so that they can return to the field, to the archive, to direct engagement with the social realities they study. The irony of the current situation is that researchers increasingly experience their subjects only through the mediation of datasets and documents, while spending their time on administrative and technical tasks that could be automated. Human-in-the-Loop reverses this pattern: machines handle the mediated, repetitive work while humans reclaim the capacity for direct encounter – for listening to the ‘silent voices’ that no dataset can fully capture. The time and energy saved through automation should be reinvested in activities that require human presence and perception: fieldwork, interviews, archival immersion, theoretical reflection.

3. The Technical Architecture: One Core, Three Repositories, Four Domains

The ontological concepts described above must be realised in concrete technical systems. The key to implementation is modularity: different capabilities are realised as independent components connected through standard interfaces, rather than bundled in a single monolithic system. This enables flexibility (components can be upgraded, replaced, or extended without affecting the whole), resilience (the failure of one component does not crash the system), and collaborative development (institutions can contribute specialised components and share them with the broader community). This section presents the methodological layer of AI4SS: the architecture that translates conceptual framework into operational infrastructure. We describe this architecture using a schematic formula – ‘one core, three repositories, four domains’ – that captures its essential structure.

The AI4SS technical architecture: One Core, Three Repositories, Four Domains

3.1 The Intelligent Runtime (‘One Core’)

At the heart of AI4SS lies what we call the Intelligent Runtime: the computational engine that drives all system operations. This runtime has two primary components.

The first is a Large Model Inference Engine providing capabilities for natural language understanding, semantic analysis, and content generation. This component leverages the remarkable advances in language modelling over the past several years, enabling systems to process unstructured text at scale, to understand the meaning of documents rather than merely matching keywords, and to generate coherent prose expressing analytical findings.

The second component is a Declarative Multi-Agent System Architecture, which we have formalised under the name POMASA (Pattern-Oriented Multi-Agent System Architecture). Rather than writing procedural code that specifies exactly how tasks should be performed, POMASA enables the definition of agent blueprints – declarative specifications of what agents should accomplish, what resources they can access, and how they should interact. These blueprints are interpreted by the runtime, which instantiates and orchestrates the specified agents to accomplish complex research tasks.

The combination of language model capabilities with multi-agent orchestration enables a qualitative leap in what AI systems can accomplish. Individual language models, for all their power, operate in isolation: they process inputs and generate outputs but cannot coordinate with other systems, cannot maintain state across interactions, cannot implement complex workflows involving multiple steps and multiple specialised functions. Multi-agent architecture overcomes these limitations by decomposing complex tasks into subtasks handled by specialised agents, coordinating the flow of information between agents, and managing the overall process from initiation to completion.

3.2 The Multimodal RAG Lake (‘Three Repositories’)

The second architectural element concerns data storage and retrieval. AI4SS operates on a multimodal Retrieval-Augmented Generation (RAG) lake: a comprehensive data infrastructure that stores research materials in forms optimised for AI processing and retrieval.

This infrastructure comprises three distinct repositories. The Structured Data Repository contains cleaned, validated quantitative datasets: statistical tables, survey results, economic indicators, and other numerical information organised in formats amenable to computational analysis. The Unstructured Materials Repository stores original texts, images, audio recordings, video footage, and other primary materials in their native formats, indexed for semantic retrieval but preserving full fidelity to the originals. The Process Assets Repository accumulates the intermediate products of research – working notes, preliminary analyses, methodological templates, best practices – enabling institutional learning and knowledge transfer.

The RAG architecture ensures that AI-generated outputs remain grounded in actual evidence. When the system formulates a response, it does not rely solely on patterns learned during training (which is the source of hallucination in standard language models); it retrieves relevant materials from the repositories and grounds its generation in this retrieved context. The result is outputs that are not only fluent but verifiable – claims that can be traced back to specific sources in the underlying knowledge base.

3.3 The Technical Capability Ecosystem (‘Four Domains’)

The third architectural element organises the diverse technical capabilities required for AI-assisted research into four functional domains, each corresponding to a distinct phase of the research process.

3.3.1 The Information Perception Domain

The Information Perception Domain handles the ingestion and preliminary processing of data. Capabilities include: web scraping and API access to open sources; integration of commercial databases and paywalled materials; multi-format document parsing (PDF, Word, Excel, PowerPoint); image understanding through multimodal models; audio and video transcription; knowledge graph construction from unstructured text; and data archiving with classification and access control. This domain corresponds to the ‘sensing’ function of the research system – its interface with the external world of information.

3.3.2 The Cognitive Reasoning Domain

The Cognitive Reasoning Domain performs deep analytical processing. Capabilities include: full-text indexing at multiple granularities (sentence, paragraph, document, corpus); deep document comprehension that grasps meaning and argument structure; research planning and pathway determination; multi-dimensional data exploration across different levels of abstraction; dynamic generation of specialised agents for specific tasks; statistical analysis and hypothesis testing; econometric methods including panel data analysis, instrumental variables, and difference-in-differences; and probabilistic reasoning for pattern recognition, semantic induction, and causal inference. This domain corresponds to the ‘thinking’ function – the analytical core of the research system.

3.3.3 The Expression Rendering Domain

The Expression Rendering Domain handles the generation and presentation of outputs. Capabilities include: report generation with logical structuring; data visualisation through charts, dashboards, and infographics; natural-language-driven visualisation generation – researchers describe what they need in a single sentence and receive a finished chart; content management with version control; multi-format typesetting for academic papers, policy briefs, presentations, articles, and videos; multi-channel distribution adaptation; data lineage tracking and citation management; and audit logging for compliance and academic integrity. This domain corresponds to the ‘speaking’ function – the system’s interface with human audiences.

3.3.4 The Collaborative Governance Domain

The Collaborative Governance Domain manages coordination and quality control. Capabilities include: team collaboration and real-time synchronisation; Git-style version control for data and documents; domain knowledge encapsulation with API exposure; cross-system protocol adaptation; citation format management (Chicago, APA, and other standards); automated fact-checking; bias detection and flagging; and workflow engines for review and publication processes. This domain corresponds to the ‘regulating’ function – the system’s mechanisms for ensuring quality, consistency, and accountability.

Together, these four domains provide a comprehensive capability architecture that supports the full lifecycle of research from data acquisition through analysis to publication and dissemination.

4. From Theory to Praxis: Applications of AI4SS

A framework is only as good as its applications. The concepts and architectures described above would remain merely theoretical were they not embodied in functioning systems that produce actual research outputs. Global South Insights has developed over forty applications within the AI4SS framework, spanning domains from political economy to agricultural research, from regional and country studies to crisis analysis. These applications run on the GSI Unified Data Platform – according to GSI internal documentation, a comprehensive data infrastructure integrating 96 datasets from institutions including the United Nations, the World Bank, and the International Monetary Fund, encompassing 41,100 indicators and 3.45 billion rows of data. This section presents four representative examples that demonstrate the viability and utility of the framework.

4.1 Meta-RAG: The Knowledge Infrastructure

Meta-RAG is the flagship implementation of AI4SS’s multimodal RAG lake concept. It is not merely a tool but an infrastructure: a comprehensive system for transforming implicit research methodologies into configurable software workflows. We describe this approach as ‘Research Methodology as Configuration’ – the idea that the procedures researchers follow can be explicitly specified, systematically implemented, and reliably repeated.

The system comprises over thirty atomic agents, including Evaluators, Orchestrators, and Validators, that can be combined into custom workflows. Users can design research processes incorporating planning, execution, and evaluation loops with human oversight at specified checkpoints. The underlying data architecture uses PostgreSQL vector databases to store multimodal knowledge bases, each containing tens of thousands of vector embeddings that enable semantic retrieval across text, images, and other materials. For any research topic, a researcher can build a specialised knowledge base containing thousands of high-quality documents in approximately one week at a cost of roughly $200 – work that would traditionally require months of literature accumulation.

Meta-RAG has been deployed to support research in political economy, agricultural technology, and Marxist studies, among other fields. A representative case demonstrates its capabilities: an African national development strategy project used Meta-RAG’s ‘New Structural Economics Lens’ workflow to produce structured policy recommendations. The system autonomously integrated macroeconomic frameworks with sector-specific industrial policies, generating a comprehensive multi-chapter report with standardised structure. Total processing time was twenty-seven minutes; total cost in computational resources was $2.91.

Such efficiency gains are significant, but they are not the primary point. What matters is that the research process remains transparent, verifiable, and reproducible. Every output can be traced to its sources; every analytical step can be examined and, if necessary, contested. The speed and economy of AI-assisted research do not come at the cost of scholarly rigour.

4.2 The Digital Sovereignty Index: Assessing Global South Autonomy

The Digital Sovereignty Index (DSI) project applies AI4SS to a substantive research question of direct relevance to the Global South: the degree to which nations control their own digital destinies. Drawing on a theoretical framework developed by our team, the DSI evaluates countries across four dimensions – data ownership, digital infrastructure, cyberspace governance, and digital capability – to produce comprehensive assessments of digital sovereignty.

The technical implementation uses the POMASA architecture with six specialised agents coordinated by an orchestrator. Agent blueprints total approximately 21,000 words; reference materials (including scoring rubrics, decision matrices, and background documentation) total approximately 188,000 words. The evaluation methodology specifies over 5,400 individual assessment criteria organised under sixteen indicators, with each indicator accompanied by multi-level scoring standards.

To date, the system has completed assessments of all eleven BRICS nations. For each country, the research process collects over 700 pieces of primary evidence, which are then deduplicated and quality-checked to yield 400–500 high-quality evidence items (a deduplication rate of 35–42%). The final output for each country is a structured assessment report of 11,000–13,000 words. Total processing time per country is approximately six hours.

External validation has been encouraging. Pablo Martinez of the National University of General Sarmiento (UNGS) in Argentina, responding to the DSI assessment of Argentina in 2024, observed: ‘Really insightful work. The capability paradox and the 2.0/5.0 score are critical findings for us.’ Such feedback suggests that AI-assisted research can produce outputs that specialists find genuinely useful – not merely plausible but substantively valuable.

4.3 The Regional and Country Studies System

The Regional and Country Studies System represents the most ambitious implementation of AI4SS to date: a comprehensive infrastructure for end-to-end academic research production. It is not an assistant that helps researchers write; it is a production system that can autonomously generate publication-ready research outputs from specified inputs and parameters.

The scale of the system reflects its ambition. It incorporates 317 declarative agents organised into 64 pipeline orchestrations. It connects to six MCP servers (providing a total of 123 tools) with access to paywalled journals, optical character recognition for scanned documents, and various analytical services. The quality assurance architecture includes 294 ‘Quality Gates’ – automated checkpoints distributed across workflow layers including literature extension, research observation, and deep analysis – that verify outputs against specified standards and flag potential hallucinations.

The efficiency gains are dramatic. Research projects that would traditionally require six to twelve months of full-time work can be completed in 40–60 hours of system runtime – an efficiency improvement of 100–200 times. Final outputs are substantial: reports of 40–80 pages (60,000–100,000 Chinese characters or equivalent), incorporating over 3,000 precisely formatted citations in Chicago or APA style.

The system has been used to produce country studies covering topics such as social governance, capital accumulation, and international dependency relations. These are not superficial overviews but deep analytical treatments that engage seriously with primary sources and secondary literature, that construct arguments and evaluate evidence, that meet the standards expected of professional research institutions.

4.4 Rapid Exploration: From Question to Insight Without Gatekeepers

The final example is not a major research programme but an impromptu exploration that illustrates a different dimension of AI4SS’s potential: the capacity for rapid, self-directed inquiry.

An Indian industry expert approached us with a question: what are the strategic contradictions in US semiconductor export controls, and what opportunities might they create for the Global South? In the traditional knowledge economy, such a question would initiate a familiar pattern of dependency: the expert would need to find someone with relevant expertise, request access to materials, wait for a literature review to be compiled, and ultimately depend on others to provide the analysis. Knowledge flows through gatekeepers; those without access to the right networks remain excluded.

Instead, we simply fed his question directly into POMASA’s generator tool. No additional information was provided beyond his original problem framing. Within thirty minutes, the system had created a complete multi-agent research apparatus – seven specialised agents implementing fourteen architectural patterns, configured to explore precisely the question he had posed. The system then ran autonomously for approximately two hours, curating 106,000 words of source material, verifying evidence, conducting analysis, and producing a twelve-chapter report of 7,600 words with 122 references.

The critical realisation is this: he could have done this himself. The entire process required no privileged access, no special expertise, no gatekeeper’s permission. POMASA is an open-source framework; the intelligent runtime environments that execute it are commercially available; the methodological patterns are documented and learnable. A researcher anywhere in the world, armed with a question and basic familiarity with the tools, can conduct the same kind of rapid exploration.

This represents a profound democratisation of exploratory research. Traditionally, the capacity to quickly survey a field, identify key debates, and generate preliminary findings has been concentrated in well-resourced institutions with extensive libraries, research assistants, and established networks. Researchers in the periphery have had to wait – for access, for translations, for someone else’s synthesis. AI4SS collapses this dependency. The cost of intellectual experimentation drops to near zero. A hypothesis can be probed in hours rather than months. A question can be explored before committing to a full research programme.

This agility does not replace rigorous, sustained inquiry; it enables it by allowing researchers to identify which questions are worth deeper attention. And crucially, it shifts the locus of control. The researcher with the question becomes the researcher with the answer – or at least with a substantive first approximation that they themselves have produced.

5. Implications for the Global South

The applications described above demonstrate that AI4SS is technically viable and practically productive. But what does this mean for the broader project of knowledge production in and for the Global South? This section considers the implications of AI-assisted social science research for scholars, institutions, and movements in the periphery and semi-periphery of the world system.

5.1 Epistemic Sovereignty

The most fundamental implication concerns what we have called epistemic sovereignty: the capacity of Global South researchers to determine their own research agendas, develop their own analytical frameworks, and control the infrastructures through which knowledge is produced and disseminated.

The current global knowledge economy is characterised by profound asymmetries. The major academic publishers are headquartered in the North; the dominant citation indices privilege Northern journals; the most powerful AI systems are developed by Northern corporations and trained on predominantly Northern data. Researchers in the Global South operate within this system but have limited influence over it. Their work is evaluated by Northern standards, published in Northern venues, and often framed in terms of Northern theoretical concerns.

AI4SS offers a path – not to autarky, which would be neither possible nor desirable, but to greater autonomy within interdependence. By building institutional knowledge bases from local materials, researchers create cognitive moats that cannot be simply appropriated or overridden. By developing analytical frameworks suited to Southern concerns – AI4SS’s theory-agnostic architecture enables the same system to operate freely across Marxist, liberal, realist, and other frameworks, so that researchers are not held captive by ideological assumptions embedded in their tools – they ensure that AI systems serve Southern agendas rather than imposing Northern categories. By maintaining human control over research direction and interpretation, they preserve the space for intellectual traditions that differ from the globalised mainstream.

This is not a matter of rejecting Northern knowledge or technology. Much of what the North has produced is genuinely valuable and should be engaged seriously. But engagement should be critical and selective, not passive and wholesale. Epistemic sovereignty means the capacity to learn from others without being subordinated to them.

5.2 Capacity Multiplication

The second implication concerns research capacity. Global South institutions typically operate with fewer resources than their Northern counterparts: smaller budgets, heavier teaching loads, less administrative support, weaker infrastructure. These constraints limit what researchers can accomplish regardless of their intellectual abilities.

AI4SS offers a form of capacity multiplication that partially compensates for resource disparities. The mechanical labour that consumes so much researcher time – literature review, data processing, formatting, quality control – can be largely automated, freeing researchers to focus on activities that require human intelligence and cannot be delegated to machines. Small teams can produce outputs that previously would have required large institutions; individual researchers can tackle questions that previously would have demanded collaborative programmes.

This multiplication is not unlimited. AI systems require infrastructure (computing resources, stable electricity, reliable internet), expertise (to design, deploy, and maintain them), and judgement (to evaluate and verify their outputs). These requirements themselves impose constraints, particularly for the most under-resourced institutions. But within those constraints, significant gains are possible. The cases presented above already demonstrate that this promise is not hollow: $200 and one week to build a specialised knowledge base, six hours to complete a comprehensive country assessment, research projects that traditionally take six to twelve months completed in 40–60 hours.

5.3 Cross-Border Collaboration

The third implication concerns collaboration across the Global South. Southern researchers have long recognised the value of South-South cooperation: the exchange of experiences, the sharing of resources, the collective development of perspectives that differ from Northern hegemonic frameworks. But such cooperation has often been hampered by practical obstacles – distance, language barriers, incompatible systems, limited funding for travel and exchange.

AI4SS provides infrastructure that can facilitate South-South collaboration at scale. Standardised protocols enable researchers using different tools and speaking different languages to work together within integrated workflows. Shared knowledge repositories allow institutions to pool their resources, avoiding duplicative effort and building cumulative knowledge bases. The GSI Unified Data Platform already offers an example: according to GSI internal documentation, the platform’s 96 datasets include 274 universal aggregation categories – BRICS nations, the African Union, groupings by income level, and more – enabling researchers from different countries to collaborate on a common data foundation. Modular architectures mean that capabilities developed in one context can be readily adapted for use in others.

The vision is one of distributed, cooperative capacity-building across the Global South. Rather than each institution developing its own tools in isolation, researchers can contribute to and draw upon a common infrastructure. Rather than competing for scarce resources, they can collaborate in mutual development. The result would be greater than the sum of its parts – a network of Southern research capacity that collectively matches or exceeds what any Northern institution can achieve alone.

5.4 Critical Vigilance

The final implication concerns the dangers as well as the opportunities of AI-assisted research. Artificial intelligence is not a neutral technology; it embeds the assumptions, biases, and priorities of those who develop and deploy it. Language models trained predominantly on English-language materials from Northern sources may reproduce Northern perspectives even when applied to Southern questions. Algorithmic systems optimised for efficiency may sacrifice nuance and context. The very power of AI tools creates risks of over-reliance – of substituting machine outputs for human judgement, of mistaking fluency for accuracy.

AI4SS responds to these dangers through the Human-in-the-Loop principle and the safeguard mechanisms that have already been engineered into practice – 294 Quality Gates, the architectural separation of content production from quality verification, end-to-end evidential traceability – these are not merely stated principles but the actual operating reality of the system. Yet mechanisms are only effective when actively wielded by critically vigilant researchers. AI systems should be treated as tools, not oracles – useful instruments that must be supervised, verified, and where necessary overridden. Human judgement must remain paramount, especially on matters involving interpretation, valuation, and normative assessment.

Critical vigilance also means attention to the political economy of AI itself. Who benefits from the development and deployment of AI systems? Whose interests do they serve? How do they redistribute power and resources? These questions should inform not only how we use AI but how we participate in its development and governance. The Global South should not be merely a consumer of technologies developed elsewhere but a participant in shaping the technological future.

6. Conclusion: Towards a Decolonised Research Paradigm

This study has presented AI4SS as a comprehensive framework for artificial intelligence in social science research, developed by and for researchers in the Global South. We have argued that contemporary social science faces structural contradictions – the bandwidth paradox, the rigour gap, the disciplinary archipelago, the labour inversion – that limit what researchers can accomplish and disadvantage those working in under-resourced institutions. We have proposed an ontological reconceptualisation through the structure of ‘One Foundation, Four Pillars, One Core’ – Trustworthy Infrastructure as the foundation, Full-Data Perception, Full-Dimensional Analysis, Full-Channel Output, and Full-Disciplinary Synergy as the four pillars, and Human-in-the-Loop as the core – that together define a new paradigm for AI-assisted research. We have described a technical architecture that realises these concepts in practice and demonstrated viability through concrete applications.

The path forward requires simultaneous development across all three layers of the framework. At the ontological level, we must continue to refine our understanding of what AI makes possible and what it threatens, how it can augment human intelligence and where it must remain subordinate to human judgement. At the methodological level, we must build and improve the technical systems that translate concepts into capabilities, ensuring that these systems embody our principles and serve our purposes. At the practical level, we must apply these systems to research questions that matter, producing knowledge that advances the interests of the peoples of the Global South.

This is not work that any single team can accomplish alone. AI4SS is offered not as a finished product but as a contribution to a collective endeavour. We invite researchers across the Global South to engage with this framework – to use it, critique it, adapt it, extend it. The tools we have developed are available; the architectures we have designed can be implemented elsewhere; the principles we have articulated can guide work in contexts quite different from our own. What we have begun, others can continue and improve.

The broader stakes are significant. The age of artificial intelligence is upon us, and its contours are still being shaped. The Global South need not be a passive recipient of technologies designed by and for the North; it can be an active participant in determining how AI develops and what purposes it serves. A decolonised research paradigm – one that harnesses technological power while preserving human agency, that learns from global knowledge while serving local needs, that builds Southern capacity while fostering global solidarity – is within reach. The task is to grasp it.

We conclude with an invitation and a challenge. To researchers in the Global South: engage with AI not as a threat to be feared or a fashion to be followed but as a capability to be mastered and directed. To institutions: invest in the infrastructure and expertise that AI-assisted research requires, and build the collaborative networks that multiply its impact. To movements: insist that technology serve the people, that the benefits of AI be broadly shared, that the age of artificial intelligence become an age of genuine human flourishing.

The future is not determined. It is made – by choices, by struggles, by collective action. AI4SS is one contribution to making a future in which the peoples of the Global South are not objects of research but its subjects, not consumers of knowledge but its producers, not followers in technological change but its pioneers. That future will not arrive on its own. It must be built. The building begins now.