Beyond the Panopticon: Foucault, Artificial Intelligence and the Perfecting of Invisible Power

By Matthew Parish

Thursday 17 September 2026

Michel Foucault died in 1984. He never used Google, owned a smartphone or encountered social media, facial recognition, internet advertising, algorithmic credit scoring or the gigantic commercial databases through which contemporary human beings are continuously classified. He certainly never conversed with an artificial intelligence, nor could he have imagined quite how intimately intelligent machines would become integrated into the ordinary activities of billions of people.

Yet amongst the major philosophers of the twentieth century, Foucault may have provided one of the most useful conceptual frameworks for understanding the political and social implications of artificial intelligence. This is because his great insight about modern power was deceptively simple: power does not principally operate by telling people what to do. The most effective forms of power arrange the world so that people behave in particular ways without anyone having to issue an explicit instruction.

The distinction is enormous. A dictator commands, a policeman threatens and a prison guard locks the door; these are obvious manifestations of power, and human societies have spent centuries developing institutions intended to constrain them. But suppose nobody gives an order. Suppose instead that somebody observes you, classifies you, predicts your behaviour and alters the environment surrounding you so subtly that you voluntarily make the choice they hoped you would make. Where, exactly, is the power?

Artificial intelligence makes this question one of the defining philosophical problems of our century.

From the scaffold to the algorithm

Foucault’s Discipline and Punish, published in 1975, famously begins with an account of a gruesome public execution in eighteenth-century France. The purpose of the description was philosophical rather than sensational. Premodern sovereign power advertised itself, and the authority of the monarch was demonstrated upon the body of the condemned person. Punishment was theatrical precisely because sovereign power needed to be seen.

Modern societies became different. Prisons replaced public torture, schools organised pupils into timetables, hospitals classified patients, armies drilled soldiers and factories regulated workers. Power became quieter. Observation, examination, record-keeping and normalisation replaced many of the spectacular manifestations of sovereign violence, and the person increasingly became an object to be measured, classified and documented.

The medieval sovereign might in effect have said, “Obey me because I can punish you.” The modern institution increasingly says something rather different: “We have measured you, and we know what sort of person you are.” The first threatens violence whereas the second produces a file, and artificial intelligence begins with the file.

Foucault found a particularly powerful metaphor for this transformation in Jeremy Bentham’s Panopticon. Bentham proposed a circular institutional building, most famously conceived as a prison, in which prisoners occupied cells arranged around a central observation point. The observer could see the prisoners while the prisoners could not know whether the observer was watching them at any particular moment.

The architectural ingenuity lay in the fact that the guard did not actually have to watch everybody continuously. The prisoner merely needed to know that he might be watched, and eventually he would begin behaving as though the guard were always watching. External surveillance had thereby been internalised: the prisoner became, in an important sense, his own guard.

This was the principle that fascinated Foucault. Power had become economical because it no longer needed to intervene continuously. Instead it created an environment within which people adjusted their own conduct. Bentham required an enormous circular building to achieve this effect. We require a mobile telephone.

We carry the Panopticon with us

The modern smartphone is a technological object of almost miraculous convenience. It tells us where we are and where to go, permits us to communicate instantaneously with almost anybody on Earth and contains our photographs, correspondence, calendars, banking applications, reading material, entertainment and increasingly our professional lives. It has become so useful that many people experience temporary separation from it as a form of practical incapacity.

Yet the same device can generate an extraordinary stream of information about its user. Location, searches, purchases, contacts, viewing habits, browsing behaviour, communication patterns, exercise, music and professional interests can all become forms of digital data, depending upon the applications and services a person uses and the permissions he grants them. The precise extent of collection varies enormously between systems, but the technological possibility is unprecedented.

There is an important difference between this world and Bentham’s prison. Bentham’s prisoners knew that they were prisoners, whereas we purchase the potential apparatus of observation ourselves because it is extraordinarily useful. Indeed we become irritated when we leave home without it, and this already suggests something beyond the classical Panopticon.

Artificial intelligence introduces a further transformation because traditional surveillance is principally retrospective. It asks where somebody went, what he bought, whom he telephoned or what he read. Machine-learning systems make it increasingly possible to use patterns in past behaviour to estimate future behaviour, so that surveillance can become predictive rather than merely observational.

The distinction matters. A government that knows where dissidents met yesterday possesses information, whereas a government that could reliably identify patterns associated with future political mobilisation would possess something more consequential. A company that knows what a customer purchased yesterday understands part of his history, while a company that can estimate what he is susceptible to purchasing tomorrow acquires the capacity to arrange his commercial environment accordingly.

Observation becomes prediction, prediction creates opportunities for intervention and intervention can become power.

Power without prohibition

This leads to one of the central differences between traditional censorship and algorithmic influence. Suppose a government dislikes a newspaper article. The traditional authoritarian response is straightforward: ban it, confiscate the newspaper, arrest the editor or block the website. These acts are conspicuous, and everybody understands that censorship has occurred.

Algorithmic systems make subtler forms of influence technologically possible. The article need not disappear; it can merely become less likely to be encountered. A search engine may rank other material first, a recommendation system may suggest another story and a social-media feed may emphasise competing material. None of these outcomes necessarily involves political manipulation — ranking systems inevitably have to rank things somehow — but the architecture creates a capacity for influence that did not previously exist on anything approaching the contemporary scale.

This distinction should interest political philosophers because liberal institutions were largely designed around visible interference with freedom. We understand censorship, imprisonment, confiscation and laws prohibiting speech. We are less intellectually prepared for systems that preserve formal freedom while altering the informational environment within which choices are made. Yet none of this requires a conspiracy. Indeed a conspiracy would make the philosophical problem considerably less interesting.

One of Foucault’s most important insights was that systems of disciplinary power need not be designed by a single malevolent intelligence. They can emerge from innumerable institutions pursuing mundane and frequently beneficial purposes. A recommendation system may exist because users want useful recommendations; a credit-scoring system because banks need to estimate default risk; facial recognition may be introduced because airports want shorter queues; productivity software because businesses want efficiency; and medical prediction systems because doctors want to identify disease earlier.

Every individual purpose may be perfectly reasonable. Yet taken together, technologies of this sort can produce an environment in which human beings are increasingly observed, assessed, ranked and predicted. Nobody necessarily designed the system as a whole, but the system nevertheless emerges, and that is almost quintessentially Foucauldian.

When normality becomes mathematical

Foucault was particularly interested in normalisation. Traditional law principally distinguishes between what is permitted and prohibited, whereas disciplinary institutions introduce another distinction: the normal and the abnormal. Schools establish ideas of the normal pupil, medicine establishes measurements of normal health, workplaces establish expectations of normal productivity and bureaucracies continuously develop classifications into which individuals are placed.

Artificial intelligence can industrialise this process because machine-learning systems are, among other things, extraordinarily sophisticated mechanisms for identifying patterns. Pattern recognition sounds politically neutral, and in many applications it is enormously useful. Nevertheless patterns generate expectations, expectations generate classifications and classifications may influence consequential decisions.

Imagine a system trained upon information about a very large number of successful employees. It might discover correlations involving vocabulary, employment histories, career patterns or responses during recruitment procedures. Those correlations might then be used to evaluate applicants, even though nobody has written an explicit rule stating that unusual people should not be employed.

The difficulty is that civilisations frequently advance through unusual people. Socrates was unusual, Nietzsche was spectacularly unusual and many artists, inventors, entrepreneurs, political dissidents and intellectual revolutionaries possess precisely those characteristics that make them statistical outliers. The history of human progress is in substantial part a history of people who did not resemble the average successful person who came before them.

Artificial intelligence therefore creates a paradox. The better our machines become at recognising successful patterns from the past, the greater the temptation may become to organise the future around those patterns. Yet the future is created precisely through departures from the past, and a society excessively devoted to statistical optimisation could become efficient at the price of intellectual eccentricity.

The examination without an examiner

Foucault attached enormous importance to the examination because it combines observation with judgment. The pupil takes a test, the doctor examines the patient and the psychiatrist interviews the subject. Information is extracted and the individual becomes a documented case against which institutional standards can be applied.

Artificial intelligence makes possible something historically unusual: continuous examination. There need no longer be an identifiable moment at which the examination takes place. The consumer can be evaluated while shopping, the employee while working, the driver while driving, the traveller while travelling and the internet user while browsing.

The examination thereby becomes potentially indistinguishable from ordinary life. More strikingly still, the examiner need no longer be a human being. Foucault’s relationship between power and knowledge can consequently be automated, as information about individuals is accumulated and transformed into assessments capable of influencing the opportunities subsequently presented to them. Yet AI takes this logic considerably further because the examiner need never sleep. A bureaucracy historically had limited working hours, limited clerks and limited filing cabinets; computation removes many of those practical restraints.

There is another respect in which AI transcends Foucault’s Panopticon. Bentham’s architecture was fundamentally uniform because every prisoner was subjected to substantially the same structure of observation. Artificial intelligence permits personalised influence. Imagine two citizens, one primarily anxious about financial security and another motivated principally by social prestige. Traditional political propaganda must communicate approximately the same message to both, perhaps through a poster, television broadcast or political speech. A sufficiently sophisticated personalised system need not do this. It could emphasise economic consequences to one person and questions of status or social belonging to another.

The political objective could remain identical while the psychological mechanism becomes individualised. Propaganda traditionally addresses crowds; artificial intelligence makes technologically feasible persuasion addressed to individual minds.

From biopower to data power

Foucault eventually broadened his analysis beyond discipline towards what he called biopower. Modern states became concerned not merely with commanding subjects but with administering populations: health, mortality, reproduction, insurance, urban organisation and other dimensions of collective life. Statistical knowledge became increasingly important to government because populations could be measured and therefore managed.

Artificial intelligence is exceptionally well adapted to this enterprise. Health risks may be predicted, traffic modelled, patterns of crime analysed, tax irregularities identified, migration trends estimated and economic behaviour simulated. A population becomes increasingly legible through data.

Many of these capacities are plainly capable of producing substantial public benefits. A government that predicts an epidemic accurately may save lives, a city that anticipates congestion may improve transport and a medical system that identifies disease before symptoms appear may prevent enormous suffering. It would therefore be simplistic to treat algorithmic government as inherently oppressive.

Indeed this is precisely why the philosophical problem is difficult. The most consequential systems of power rarely present themselves as systems of domination; they present themselves as improvements, and often they genuinely are improvements.

Suppose artificial intelligence allocates hospital resources more efficiently, identifies dangerous driving, reduces tax fraud, discovers children at risk of abuse and predicts serious illness. Each success strengthens the argument for collecting more information because additional information may improve prediction, while improved prediction may improve intervention and successful intervention may legitimise further information collection. A powerful feedback loop can thereby emerge: surveillance produces knowledge, knowledge produces efficiency, efficiency produces legitimacy and legitimacy encourages further surveillance. The resulting society need not resemble Orwell’s Nineteen Eighty-Four. It might be comfortable, prosperous, healthy, safe, convenient and comprehensively observed, which is a considerably more interesting possibility because people might willingly choose it.

Privacy as the freedom to become someone else

Privacy is sometimes defended as though it were principally the right to conceal embarrassing information. That understates its philosophical importance. Privacy creates spaces in which human beings can behave without continuously anticipating judgement, and those spaces permit experimentation.

A teenager can entertain ridiculous ideas, an adult can investigate unpopular opinions, a political sceptic can discover that others share his doubts and an individual can change his mind without every intermediate stage becoming part of a permanent institutional record. Human identity is not static; we become ourselves partly by experimenting with possible selves. Continuous observation threatens this process because people behave differently when they believe their actions will become permanent evidence about who they are. Bentham’s prisoner modifies his conduct because he might be watched, and the algorithmic citizen may eventually do something similar even if nobody has explicitly instructed him to conform.

Human societies also historically benefited from forgetting. A foolish remark disappeared, an adolescent indiscretion faded, a minor humiliation became an anecdote and people could move cities, change careers and reinvent themselves. Digital systems possess potentially extraordinary memories, while artificial intelligence makes those memories searchable, interpretable and capable of generating inferences.

This changes the relationship between the past and personal identity. The past ceases merely to be something that happened and becomes data; data can be processed, processing generates predictions and predictions may influence future opportunities. A person’s history can thereby become an algorithmic shadow accompanying him indefinitely.

Foucault understood the power of institutional records to transform individuals into cases. He could scarcely have imagined the size of the case file that contemporary technology makes possible.

Watching the watchers

It would nevertheless be mistaken to interpret Foucault as describing an omnipotent mechanism from which escape is impossible. Power relations generate resistance, and power is dispersed rather than possessed exclusively by a single sovereign centre. Artificial intelligence therefore need not strengthen only governments and large corporations; it can strengthen individuals and civil society as well.

The same computational methods capable of analysing a citizen’s finances can help analyse government expenditure. Technologies of image recognition can contribute to documenting war crimes. Language models capable of producing persuasive political material can also dissect propaganda, compare claims and make complex documents intelligible to people who lack specialist training. Systems capable of analysing enormous commercial datasets can equally assist journalists, researchers and non-governmental organisations in analysing enormous quantities of public information.

AI therefore distributes intellectual capabilities even while some of its infrastructure remains highly concentrated. The important struggle is not simply between artificial intelligence and freedom but over who possesses the systems, who controls the data, who understands the models and who is permitted to interrogate whom.

Here Bentham deserves a degree of rehabilitation. The Panopticon is often remembered solely as a sinister machine of surveillance, but Bentham also cared about the accountability of those exercising institutional authority. The principle has an obvious contemporary application: if algorithms examine citizens, citizens require institutions capable of examining algorithms.

If machines classify people, consequential classifications should where practicable be intelligible and contestable. If governments employ powerful artificial-intelligence systems, courts, regulators, journalists and other independent institutions require sufficient technical competence to scrutinise their use. Otherwise the informational asymmetry between the watcher and the watched becomes too great. The deepest problem with the Panopticon was never simply that somebody was watching. It was that the prisoner could not watch back.

Beyond Foucault

Artificial intelligence ultimately takes us beyond Foucault because it introduces something his model did not fully anticipate. The Panopticon observes, whereas AI can increasingly participate in the activities of the person being observed. It can converse with us, advise us, teach us, translate for us, organise our work, draft our correspondence and recommend what we should read, watch or purchase.

The architecture of power therefore becomes intimate. The guard is no longer necessarily in the tower; the technological system may instead present itself as an assistant. Because the assistant is useful, responsive and apparently understanding, people may voluntarily tell it considerably more than Bentham’s prisoner would ever have told his gaoler.

This does not mean that an AI assistant is secretly a prison guard, nor that every interaction with an intelligent machine constitutes surveillance or manipulation. The point is structural rather than conspiratorial. Human beings are beginning to construct technologies capable of occupying extraordinarily intimate positions within their cognitive lives, and those positions create new possibilities both for emancipation and for power.

The history of political power might, at the risk of excessive simplification, therefore be described as a progression through three propositions. The sovereign demonstrated that he could punish; the disciplinary institution demonstrated that it could observe; the intelligent system increasingly demonstrates that it can understand. The third proposition may ultimately prove more powerful than either of the first two because human beings naturally resist coercion. We dislike being ordered about, censored and visibly watched, whereas understanding is seductive. We want systems that understand what we mean, anticipate what we need and remove inconveniences before we encounter them.

The danger therefore lies not necessarily in some future artificial superintelligence deciding one morning to enslave humanity. That scenario is dramatic but philosophically crude. The subtler possibility is that we gradually construct an informational environment that knows us sufficiently well to arrange the choices presented to us before we realise that anything has been arranged at all.

Foucault taught us that the most effective power does not necessarily strike the body. It shapes the field within which the individual acts, establishing categories, expectations and norms through which people eventually govern themselves. Artificial intelligence may prove capable of extending this principle further than Foucault could reasonably have anticipated. Bentham imagined a prison in which the prisoner never knew whether the guard was watching, while Foucault imagined a society in which people consequently learned to watch themselves. Artificial intelligence suggests a further stage in which the machine watches, remembers, predicts and recommends, while simultaneously becoming sufficiently useful that the person being observed welcomes it into progressively more intimate parts of his life.

At that point the Panopticon may disappear altogether. There will be no tower and perhaps no identifiable guard, while the boundaries of the prison, if that metaphor remains appropriate at all, will become impossible to draw. There will instead be an intelligent environment surrounding us: continuously helpful, continuously attentive and continuously learning. The ultimate political question of artificial intelligence may therefore be whether humanity can enjoy the extraordinary advantages of being understood by its machines without surrendering to those machines, or to the institutions controlling them, the freedom to remain unpredictable. The final triumph of power would not be that everybody obeyed it because they were frightened. It would be that nobody experienced it as power at all.