(I wrote this draft a year ago. August of 2025. Then my ADHD brain forgot to publish it).
A small heads up. This article is probably sitting on the top of the Dunning-Kruger curve. However, I still think the article holds some water. It is just the documenation of my train of thoughts. Having said that, let’s start.
One of my favourite scenes from Game of Thrones is this one. It is a self-contained scene and does not contain any spoiler, so you can happily proceed if you still haven’t watched this show.
In this scene, the character (male) Littlefinger said – “Knowledge is power” and, was boasting how he used knowledge to tame the power-based hierarchy in the society throughout his life. Then the character Cersei (female, visibly representing a upper hierarchy) overpowered him to prove the point that “power is power” – painting an omnipresent, knowledge-agnostic view of power.
Power is power – Cersei Lannister
However, Michael Foucault (1926-1984) would completely disagree with both of them. To understand why – we need to understand the basics of Power-Knowledge construct.
Power-Knowledge Construct
Let me describe the idea of Foucault’s Power-Knowledge construct first, because it is central to this article. Let’s start with a level 101, stripped off definition.
As per Foucaults, power and knowledge are not separate entities. They can not exist independently in isolation, rather they create a combined construct. Those with knowledge gain power, and those with power decide what counts as knowledge – or what type of knowledge to be purused.. Throughout history, whoever held power determined what counted as “real” knowledge and what didn’t. In medieval times, the Church had the power and declared certain religious texts as truth while dismissing other ideas as heresy.
The oversimplication above downplay the depth of this idea. So below is a more objective definition – which I still think is accessible.
Power-knowledge operates through “discourses”—entire systems of thought, language, and practice that define what can be said, thought, or done in a given field.
Critically, power-knowledge doesn’t just control people from the outside. Rather, it produces “subjects” via which it shapes how people understand themselves and govern their own behavior. Modern sexuality is an example of this. The entire discourse of sexuality is a power-knowledge apparatus that makes people internalize certain norms and monitor themselves. This power doesn’t flow from a sovereign or ruling class downward; instead, it circulates through networks at every level of society in what Foucault called the “microphysics of power.”
Large Language Models as Power-Knowledge Artifacts
Large Language Models (LLM) are the hallmarks of power-knowledge construct. That’s not only because its existence itself is a grand display of power (which is required for the accumulation of compute and the capital to sustain them). It is specially because it operates simultaneously as products of existing power structures and instruments that reproduce/amplify them.
Full disclaimer: it is not an Us vs Them narrative, rather an empirical observation. Following are some ways the power-knowledge manifested itself in building the LLMs.
The Training Corpus and Tokenization
This one is obvious. Needless to say – the frontier models are trained predominantly on English content, with heavy representation of Western sources. It is not because of conscious bias. West has dominated digital content production, academic publishing, and internet infrastructure for decades. It reflects historical patterns of colonial power, economic dominance, and linguistic hegemony.
The very first step of the training, which is known as Tokenization has been shown to compress information very inefficiently for a large class of languages. Which means – from information theory perspective – the same number of tokens can contain significantly less amount of information for a language with non-Germanic/non-Latin origin under the prominent tokenization processes. The training corpus + the tokenization process that’s optimized for certain languages lead to asymmetric representation of the languages – and the ideas those are sometimes exclusive to those languages/underlying language generation processes..
The Flattening of Non-Western Epistemologies
Mostly due to the above – Large language models lack robust representation of non-Western epistemologies. They can aptly describe Ubuntu philosophy, Confucian ethics, or Indigenous knowledge systems, but they describe them through a Western explanatory framework. They translate these traditions into something legible to their dominant paradigm, potentially flattening their internal logic in the process.
Dictating the Conditions of Knowledge Creation
One of the biggest contributions of AI right now is in the field of Protein Folding problem. Moreover, AI is now being extenively used for material discovery. So, AI is not only disrupting white-collar knowledge works – which really do not create any knowledge, rather distribute it under some form of bueraucracy. AI is also paving way for the creation of new knowledge frontiers. This is where the power-knowledge dynamics become a little tricky because it makes us asking the question – how unbiased this knowledge discovery process is?
The Mathematical Parallel
In mathematics, AI is being used for “autoformalization”—automatically translating mathematical proofs into formal languages that computers can verify. The goal is to use large language models to discover new theorems and conjectures at scale. This approach works partly because mathematical truth has relatively well-defined syntactic structures and proof verification procedures. Even there, the choice of foundational systems (ZFC set theory, type theory, etc.) isn’t philosophically neutral, but there’s some convergence on what constitutes valid reasoning.
The Philosophical Divergence
Now consider applying this approach to philosophy: using large language models to formalize philosophical frameworks and automatically generate new philosophical arguments at scale. This represents knowledge creation “one degree above”—not just reproducing existing knowledge but proposing new philosophical arguments and conjectures.
But philosophy has no such convergence. There isn’t even agreement on what makes a philosophical argument good, let alone valid. Western analytic philosophy values logical consistency, clear definitions, and argumentative rigor in a specific style. Continental philosophy values different things—genealogical depth, rhetorical power, attention to what’s excluded by formalization itself. Non-Western traditions might prioritize harmony, paradox, narrative, or practical wisdom over systematic argumentation.
Scenario 1: Reinforcement Through Formalization
If philosophical frameworks are formalized using large language models trained on predominantly Western corpora, several concerning patterns emerge:
Structural privilege: Arguments that fit propositional logic, premise-conclusion structures, and analytic philosophy’s style will formalize “cleanly.” Arguments that work through metaphor, aphorism, dialogue, or contextual reasoning will seem fuzzy, informal, or deficient—not because they lack rigor, but because they lack this particular kind of rigor.
Ontological bias: The very categories used for formalization—belief, proposition, agent, rationality—come loaded with Western metaphysical assumptions. Formalizing Buddhist philosophy of no-self, or Daoist notions of wu-wei, might require forcing them into conceptual schemes that fundamentally distort their meaning. The framework itself becomes a filter that only certain philosophical traditions can pass through intact.
Citation loops: Large language models generating new philosophical conjectures will likely recombine and extend arguments from their training distribution—which means Western philosophers citing Western philosophers, creating an exponentially reinforcing feedback loop. Each generation of “new” philosophy is built on the same foundations, extending the same traditions.
Scenario 2: The Flattening of Philosophy
Perhaps worse than dominance is homogenization. If large language models generate philosophy at scale, we might see:
- Philosophy optimized for formalizability rather than insight
- A narrowing of what counts as “legitimate” philosophical method
- The erosion of traditions that don’t translate well into the kinds of structures large language models can manipulate
- A new regime of truth where “good philosophy” means “amenable to computational verification”
This would be power-knowledge operating at the level of epistemology itself: not just controlling what we think, but howthought must be structured to count as knowledge.
The Recursive Amplification Problem
This creates a vicious cycle of increasing entrenchment:
- Large language models trained on Western-dominant corpora
- Used to formalize philosophical frameworks (privileging Western structures)
- Generate “new” conjectures (recombinations within Western paradigms)
- These get published, cited, become part of the training data
- Next generation of models has even stronger Western bias
- Formalization systems become even more deeply entrenched in Western categories
Each iteration naturalizes the dominance more deeply. The system doesn’t just reproduce Western philosophy—it makes Western philosophical methods seem like the inevitable structure of philosophical thought itself. What was once a contingent historical development becomes encoded as the necessary form of reason.
Conclusion: The Unresolved Question
The prospect of AI-generated philosophy at scale forces us to confront how power-knowledge structures don’t just shape what we know, but the very mechanisms through which knowledge comes into being. Large language models, trained on corpora that reflect centuries of unequal power relations, now stand poised to become not just repositories of knowledge but generators of it.
The question remains open: Can systems for AI-generated philosophy be built that don’t collapse into recursive reinforcement of existing hegemonies? Or is this challenge inherent to the project itself—a manifestation of power-knowledge so deep that it structures the very idea of what it means to create philosophical knowledge?
Perhaps the most important intervention is simply to make these dynamics explicit, to denaturalize the apparent inevitability of these systems, and to insist on asking: Whose philosophy? Whose methods? Whose knowledge? And ultimately: Who benefits from determining the answers?
Shah Mahdi Hasan 