Artificial Intelligence and the Enlightenment: Foucault’s Unfinished Question

By Matthew Parish

Saturday 29 August 2026

There is a curious temptation, whenever humanity invents something extraordinary, to imagine that it has invented something entirely new. Artificial intelligence is currently suffering from this treatment. It is variously described as the end of work, the beginning of machine consciousness, the destruction of education, the democratisation of knowledge, the abolition of truth and the first step towards the replacement of humanity. Much of this language tells us less about artificial intelligence than it does about ourselves. Every age projects its anxieties upon its most powerful technologies.

A more interesting way of thinking about artificial intelligence is to place it within a much older intellectual project — the Enlightenment. AI may be understood as one of the most remarkable products of the Enlightenment ambition: the attempt to organise knowledge rationally, to make information universally accessible, to submit inherited authority to criticism and to enlarge the capacity of ordinary human beings to reason for themselves.

Yet there is an obvious complication. The Enlightenment was never merely about accumulating knowledge. It was about emancipation. And a machine that answers our questions for us might conceivably liberate us from ignorance while simultaneously relieving us of the burden of thinking.

This paradox brings us to one of the most provocative twentieth-century discussions of the Enlightenment: Michel Foucault’s 1984 essay Qu’est-ce que les Lumières? — What is Enlightenment?

Kant’s question

Foucault began with an essay written exactly two centuries earlier. In 1784 Immanuel Kant answered the question Was ist Aufklärung? — what is Enlightenment? — by describing Enlightenment as humanity’s emergence from a condition of self-incurred immaturity.

Kant’s famous injunction was Sapere aude: dare to know. Have the courage to use your own understanding.

The important word is own.

For Kant, ignorance was not merely the absence of information. Human beings remained intellectually immature because they permitted other people to think for them. Priests supplied their beliefs, physicians prescribed their habits and political authorities determined the boundaries of permissible thought. Enlightenment therefore required an exercise of intellectual courage. One had to accept responsibility for one’s own reason.

This immediately presents an uncomfortable question about artificial intelligence.

Suppose that virtually every educated person possesses an intellectual companion capable of explaining almost any subject, translating almost any language, summarising enormous bodies of literature, drafting arguments, testing hypotheses, writing computer programs and identifying weaknesses in reasoning. Has humanity thereby achieved an unprecedented extension of the Enlightenment?

Or has it created the most sophisticated guardian Kant could ever have imagined — an authority so convenient that people gradually cease to exercise their own understanding?

The answer may be both.

Foucault’s Enlightenment

Foucault’s treatment of Kant is interesting precisely because Foucault was not conventionally an Enlightenment philosopher. His writings repeatedly investigated the ways in which institutions that purported to embody rational progress — prisons, hospitals, psychiatric institutions, systems of sexuality and administrative bureaucracies — simultaneously created systems of classification, discipline and power.

Modernity did not simply liberate people from medieval superstition. It invented new ways of governing them.

Yet Foucault did not simply reject the Enlightenment. In What is Enlightenment? he suggested something subtler. Enlightenment should not necessarily be understood as a collection of eighteenth-century doctrines to which modern people must swear allegiance. It might instead be understood as an attitude — a continuing critical interrogation of the present and of ourselves.

That is a remarkably fruitful way of thinking about AI.

The relevant question is not whether artificial intelligence is an Enlightenment technology. In an obvious historical sense it is. Computers, formal logic, statistics, mathematical modelling, experimental science and the enormous accumulation of recorded human knowledge upon which large language models depend are all descendants of the rationalising intellectual traditions associated with modernity.

The more Foucauldian question is: what kinds of human beings will artificial intelligence cause us to become?

The universal library

The optimistic answer is extraordinary.

For most of human history knowledge has been scarce. Books were expensive. Libraries were geographically concentrated. Universities were available to tiny minorities. Languages divided intellectual communities. Specialist expertise was difficult to acquire and even more difficult to navigate.

The internet transformed this situation but only incompletely. It created an immense library without creating a universally competent librarian.

Artificial intelligence begins to supply the missing librarian.

A student in Lviv can ask for an explanation of Kant’s transcendental deduction. A farmer in Kenya can ask questions about soil chemistry. A programmer in Bangalore can learn an unfamiliar computer language. A pensioner in Yorkshire can explore the history of the Peloponnesian War. A Ukrainian refugee can draft a letter in German. A schoolchild can ask why the stars shine.

The cost of obtaining an intelligible first explanation of almost anything is approaching zero.

This is an Enlightenment achievement of almost unimaginable magnitude.

Voltaire, Diderot and the Encyclopédistes dreamed of collecting human knowledge and making it systematically available. Their Encyclopédie eventually contained tens of thousands of articles and represented one of the great intellectual achievements of eighteenth-century Europe. Yet even the most ambitious Encyclopédiste could scarcely have imagined a system capable of conducting an individual conversation with millions of readers simultaneously, changing the level and style of explanation for each of them.

AI therefore carries the encyclopaedic ideal towards its logical extreme.

But Foucault teaches us to become suspicious precisely at the moment when a system announces that knowledge has become universal.

Knowledge and power

One of Foucault’s enduring insights was that knowledge and power cannot always conveniently be separated.

Systems of knowledge classify. Classification creates norms. Norms distinguish the normal from the abnormal, the sensible from the irrational and sometimes the permissible from the impermissible.

Artificial intelligence necessarily does these things.

An AI system is not a transparent window onto the totality of human knowledge. It is a constructed artefact. Its training materials have been selected through countless historical accidents and institutional decisions. Its behaviour is subsequently modified through engineering, evaluation, policy and legal requirements. Its answers are shaped by statistical patterns within its training and by decisions about how it ought to interact with people.

None of this entails conspiracy. Indeed that is precisely why Foucault remains useful. His conception of power was generally more sophisticated than the proposition that somebody maliciously sits in a room deciding what everyone else must think. Power can arise from systems, classifications, professional conventions and institutional practices without possessing a single conspiratorial author.

AI systems may therefore become instruments of what we might call epistemic normalisation.

If hundreds of millions of people begin asking machines what is reasonable, true, appropriate, offensive, historically accurate or morally defensible, then the assumptions embedded within those machines acquire enormous cultural significance.

The danger is not principally that AI will order people what to think. It is that people will ask it.

The new pastor

Kant’s immature person has a book that understands for him, a pastor who provides his conscience and a physician who determines his diet. Once such guardians are available, Kant observed, thinking for oneself appears difficult and dangerous.

The analogy with AI is almost embarrassingly precise. Why struggle through a difficult book when the machine can summarise it? Why formulate an argument when the machine can draft it? Why learn another language when simultaneous translation is becoming effortless? Why remember historical facts when they can be retrieved instantly? Eventually one might ask an even more disturbing question: why decide what one thinks when the machine can formulate a plausible opinion?

Here the Enlightenment machine threatens to become an anti-Enlightenment device.

The problem is not that the machine thinks. The philosophical question of whether contemporary AI genuinely thinks may ultimately prove less important than another question: does the human being still think when using it?

A calculator can either extend mathematical competence or destroy elementary numeracy, depending upon how it is used. Search engines can enlarge curiosity or replace memory. Social media can democratise publication or create mechanisms of conformity.

AI magnifies this familiar ambiguity because it does not merely retrieve information. It participates in reasoning expressed through language. Language is where much of human reflection takes place.

Criticism as a way of life

This is where Foucault’s reinterpretation of Enlightenment becomes especially valuable.

If Enlightenment is understood simply as the accumulation and dissemination of knowledge, then AI represents its triumph. Humanity has created machines capable of manipulating quantities of recorded information beyond the capacities of any individual human mind.

But if Enlightenment is an ethos of criticism, the conclusion becomes less straightforward.

For Foucault, the Enlightenment attitude involved a critical investigation of the limits imposed upon us. We inherit categories, institutions, assumptions and forms of identity and then ask how necessary they really are.

Artificial intelligence can participate magnificently in this exercise.

One can ask an AI system to attack one’s argument rather than confirm it. One can ask it to reconstruct the strongest possible case for an opposing position. One can compare intellectual traditions that would previously have required years of specialist reading. One can test whether apparently obvious assumptions survive confrontation with historical examples. One can interrogate one’s own prejudices.

Used this way, AI is not an oracle. It is a philosophical sparring partner. And that distinction may prove fundamental.

The worst intellectual relationship with artificial intelligence is passive consumption. The machine provides an answer and the human being accepts it because the prose sounds authoritative.

The better relationship is adversarial cooperation. Why do you think that? What evidence supports this proposition? What is the strongest objection? What have you omitted? Are you confusing correlation with causation? What would Kant say? What would Nietzsche say? What would Foucault say? Now tell me why Foucault might be wrong.

Such exchanges would have been astonishing luxuries to earlier generations of scholars. AI can therefore make intellectual criticism cheaper. Whether it makes human beings more critical remains another matter.

The Enlightenment and its shadows

Foucault also reminds us why naïve technological optimism is inadequate.

The Enlightenment produced extraordinary human achievements: constitutional government, modern science, religious toleration, expanding conceptions of individual liberty and eventually universalist doctrines of human rights. Yet modern rationality also accompanied empire, racial classification, industrial warfare, bureaucratic surveillance and technologies of social control.

Reason does not guarantee benevolence. Artificial intelligence inherits the same ambiguity.

The capacity to analyse enormous quantities of information may help diagnose diseases, discover new materials and improve public administration. The same capacity can facilitate surveillance. Generative technologies may democratise artistic production while making forgery effortless. Language models may broaden education while enabling industrialised propaganda. Autonomous systems may save human labour while transforming warfare.

There is no contradiction here.

Technologies amplify capacities. They do not determine the purposes towards which those capacities are directed. The Enlightenment mistake — insofar as there was one — was sometimes to imagine that increases in rational capacity necessarily generated moral progress. The twentieth century disposed brutally of that assumption. AI should dispose of it once again.

Who constructs the truth machine?

There is also a political economy of artificial intelligence. The Encyclopédie required publishers and patrons. Universities require governments, donors and students. Newspapers require proprietors and readers. AI requires immense quantities of computing power, electricity, data, capital and technical expertise.

The universal intellectual companion therefore emerges from some of the most concentrated centres of technological and financial power in human history. A Foucauldian analysis cannot ignore this.

Who trains the models? Who defines acceptable behaviour? Who controls access? Which languages are richly represented and which are marginal? Which societies generate the digital records from which models learn? Which legal systems determine what they may say? Which governments can compel access to their infrastructure? Which commercial incentives determine their development?

These questions do not demonstrate that AI is illegitimate. They demonstrate that AI is political in the broadest sense. Every library has an architecture. Every archive contains some things and excludes others. Every educational institution has a curriculum. Every AI model likewise represents a particular organisation of knowledge. The proper Enlightenment response is neither blind confidence nor paranoid rejection. It is criticism.

The courage to use somebody else’s intelligence

There is nevertheless something peculiar about applying Kant’s injunction to AI. Sapere aude traditionally means having the courage to use one’s own understanding. Yet human intelligence has never been entirely individual. We think using languages we did not invent. We inherit concepts from dead philosophers. We rely upon scientific discoveries we cannot personally verify. We consult colleagues, teachers, encyclopaedias and libraries.

Human reason is already collective. AI merely makes this fact spectacularly visible.

The relevant distinction may therefore not be between thinking independently and accepting external assistance. Nobody thinks independently in that absolute sense.The distinction is between delegating cognitive labour and delegating intellectual responsibility.

We may ask a machine to find arguments. We should remain responsible for deciding whether they are good. We may ask it to summarise evidence. We should remain responsible for deciding whether the evidence is sufficient. We may ask it to draft our words. We should remain responsible for deciding whether those words express what we actually believe. This principle might constitute an Enlightenment ethics for artificial intelligence. Use every intellectual tool available. Surrender judgment to none of them.

AI and the creation of the self

Foucault’s later philosophy increasingly concerned what he called technologies of the self — practices through which individuals constitute themselves as particular sorts of subjects.

Artificial intelligence may become the most intimate technology of the self ever devised.

People already use AI to write diaries, examine relationships, consider careers, explore political beliefs, learn subjects, compose correspondence and organise their lives. Future systems may know an individual’s writings, preferences and intellectual history sufficiently well to conduct extraordinarily sophisticated conversations about that person’s developing ideas.

This could encourage unprecedented self-examination. It could equally create unprecedented dependence. The danger is not necessarily some cinematic machine rebellion. A subtler danger would be a generation of people who become extraordinarily capable while becoming uncertain where their own intellectual agency begins. They might write beautifully without mastering prose, argue persuasively without understanding logic and possess immediate access to virtually unlimited knowledge without developing much knowledge of their own. They would appear spectacularly enlightened while remaining, in Kant’s sense, curiously immature.

A new Enlightenment

The appropriate response is not to prohibit AI or romanticise a pre-digital intellectual world. Every transformative technology generates nostalgia for the cognitive habits it destroys. Socrates famously worried that writing would weaken memory. He was correct. Writing nevertheless made philosophy, history, modern science and complex civilisation possible. AI will undoubtedly weaken some intellectual capacities because human beings cease practising skills that machines perform better.

The question is which capacities we must preserve. Memory may become somewhat less important. Routine composition may become less important. Mechanical research may become less important. Judgment will become more important. So will curiosity, scepticism, conceptual clarity, moral imagination and the ability to recognise when an apparently convincing argument rests upon nonsense.

The education appropriate to an AI civilisation should therefore become more philosophical rather than less. Students should be taught not merely to obtain answers but to interrogate them. They should learn logic, epistemology, rhetoric, history and the philosophy of science. They should understand how claims acquire authority and how apparently neutral systems embed assumptions. In other words, the arrival of machines capable of generating knowledge-like language makes the Enlightenment attitude more necessary than ever.

What is Enlightenment — now?

Foucault’s great insight was that Kant’s question should never finally be answered.

What is Enlightenment? is not merely a historical question about eighteenth-century Europe. It is a question each generation asks about its own relationship with authority, knowledge and freedom.

Our generation happens to be asking it while constructing machines capable of speaking with the accumulated voices of civilisation. That is both exhilarating and dangerous.

Artificial intelligence could become the greatest Enlightenment instrument humanity has created — a universal tutor, translator, research assistant and intellectual interlocutor available eventually to almost every human being.

It could also become the perfect technology of intellectual dependency: an omnipresent authority consulted precisely because consulting it is easier than thinking. Foucault would presumably have resisted telling us which destiny awaits us. Technologies do not arrive with predetermined political meanings. Their significance emerges through the practices, institutions and forms of human behaviour that grow around them.

The decisive issue therefore lies not inside the machine. It lies in the habits of the person sitting before it. The Enlightenment was never ultimately a project to create a world containing more information. It was a project — imperfect, contradictory and perpetually unfinished — concerned with humanity’s relationship to reason and authority.

Artificial intelligence has suddenly made that relationship visible again. We have constructed something that can answer us. The Enlightenment question is whether we will continue to question it.