Sociology of, and in, the age of AI

§ 1Why sociology, specifically

Public debate about AI is dominated by three professions. Engineers ask whether a system works. Economists ask what it will do to productivity and jobs. Ethicists ask whether its use is right. All three questions matter. But each tends to treat society as a backdrop: a market the technology enters, a population it affects, a set of values it should respect.

Sociology starts from the other end. It asks how a technology is made of social relations, how it is taken up by institutions with their own interests and habits, and how it changes the relations between people once it is there. Four observations make this concrete.

1. The data is society, recorded. A language model is trained on the written record of human life, with all its inequalities of who got to write, in which language, and about whom. A hiring model is trained on past hiring decisions. In both cases the model learns social structure, and then reproduces it with the authority of a machine.
2. The labour is hidden. Behind every “automatic” system are people labelling images, rating answers, moderating content, often on piece rates, often in the Global South.
3. The effects run through institutions. An identical model means different things in a hospital, a police department, a school and a welfare office, because each institution decides what to measure, who can appeal, and who bears the cost of errors.
4. People respond. Classified people adjust to being classified. Workers game metrics. Users anthropomorphise chatbots. These responses change the systems in turn.

The sociologist of science Sheila Jasanoff's term for this is co-production: technology and social order are made together, each shaping the other. Nothing in this section says AI is good or bad. It says AI is social, and that you cannot understand it otherwise.

§ 2Six lenses

Each classical tradition gives a different picture of the same technology. They are not rival answers so much as different instruments, each showing something the others miss.

A few elaborations are worth making.

Weber's cage, automated. Weber admired bureaucracy's efficiency and feared that it would reduce people to cogs and escape political control. An algorithmic system takes both features further. It applies rules more consistently than any clerk, and it can be more opaque than any clerk: there is often no one in the room who can explain why the decision went the way it did. The sociologist Jenna Burrell (2016) distinguished three kinds of opacity — corporate secrecy, technical illiteracy, and the inherent complexity of machine learning models — and only the first can be solved by disclosure. The organisational scholar Hatim Rahman, studying a large freelancing platform, described workers subject to an opaque rating algorithm whose rules kept changing as living in an “invisible cage” (2021).

Du Bois's double consciousness, quantified. To live as a target of classification is to learn to see yourself through its eyes: to wonder what your credit score says about you, to shape your CV for the screening software, to change how you speak to pass the voice analysis. Ruha Benjamin's Race After Technology (2019) named the result the “New Jim Code”: discriminatory designs that encode inequity while appearing neutral, objective or even progressive.

Goffman's audiences. People now routinely manage impressions for non-human audiences. Job applicants are coached to include keywords for applicant-tracking systems and to perform for video interviews scored by software. Creators speak of “the algorithm” as a capricious audience to be pleased, a form of what Taina Bucher (2017) called the algorithmic imaginary: the ways people imagine and feel about algorithms, which shape how they behave regardless of whether the imagined model is accurate.

§ 3Classification and inequality

Sorting people is one of the oldest functions of the state, as the census, the passport and the IQ test remind us. Machine learning is, at its core, a technology for sorting at scale. Marion Fourcade and Kieran Healy's The Ordinal Society (2024) argues that this has produced a new social order, in which life chances depend increasingly on rankings derived from digital traces: credit scores, risk scores, ratings, engagement metrics. Where Weber spoke of class situations defined by market position, Fourcade and Healy speak of classification situations, defined by position in these rankings.

How does inequality get into a model? Sociologists and computer scientists have identified several routes, which are worth separating because the remedies differ:

• Historical bias. The world the data records is unequal. A model that accurately learns from past hiring will learn past discrimination. Amazon reportedly abandoned an experimental recruiting tool around 2018 after finding it penalised CVs containing the word “women's”, because it had learned from a decade of mostly male hires.
• Representation bias. Some groups are under-represented in the data, so the model works worse for them. This was the Gender Shades finding (see the studies).
• Measurement bias. The thing measured is a poor proxy for the thing that matters, and the gap differs by group. The hospital algorithm used cost for need (see the studies). Arrests are used for crime.
• Deployment bias. A system is used in a context or for a purpose it was not built for, or its outputs are interpreted differently for different people.

When the state automates

Virginia Eubanks's Automating Inequality (2018) studied three American cases: an automated welfare eligibility system in Indiana that denied a million applications in its first three years, often for minor paperwork errors; a coordinated entry system for homeless services in Los Angeles; and a child-maltreatment screening model in Allegheny County, Pennsylvania. Her argument was that such systems build a “digital poorhouse”: the poor are subjected to intense data collection and automated suspicion that wealthier citizens never experience.

Three government scandals have since shown what this looks like at national scale:

The pattern across all three is sociological rather than technical. Each system was built to save money and catch cheats; each was deployed against people with little power to contest it; each reversed the presumption of innocence; and each was corrected only after sustained public, legal and political pressure. The A-level case is notable for one more reason: it was the moment young people in Britain discovered that a model, not a person, had decided their futures, and revolted against it.

Search, representation and harm

Safiya Umoja Noble's Algorithms of Oppression (2018) began with a simple observation: a search for “black girls” returned pornographic results. She argued that commercial search engines, which present themselves as neutral windows on the web, in fact reflect and amplify racist and sexist representations because they are built around advertising and popularity. Generative image and language models have repeatedly shown similar patterns: prompts for professionals that default to white men, for criminals that default to darker skin. Developers have worked to reduce these, and the corrections themselves have sometimes produced new controversies, which illustrates that there is no neutral setting: every choice about how a model should represent society is a choice about what society should look like.

§ 4Prophecies that fulfil themselves

The Thomas theorem (1928) said situations defined as real are real in their consequences. Merton's self-fulfilling prophecy (1948) said a false belief can produce the behaviour that makes it true. Machine learning systems that inform decisions are prediction machines, and when predictions guide action, they can produce their own confirmation.

Predictive policing is the clearest case. If a model sends more officers to a neighbourhood, they will record more crime there — not necessarily because more happens, but because more is seen. That data then trains the next model, which sends officers back. Kristian Lum and William Isaac (2016) simulated this using a widely used predictive policing algorithm and drug-crime data from Oakland, California. Public health survey data suggested drug use was spread fairly evenly across the city, but the algorithm, trained on arrest records, would have concentrated policing in poor, largely Black and Latino neighbourhoods, generating more arrests there and reinforcing itself. Computer scientists later formalised the problem as a “runaway feedback loop”.

Sarah Brayne's Predict and Surveil (2020), based on years of fieldwork inside the Los Angeles Police Department, showed what this looks like in practice: officers mixing algorithmic predictions with their own judgement, the expansion of surveillance to people with no police contact through network data, and resistance among officers who felt the systems second-guessed their expertise.

The same logic appears elsewhere. A credit model that refuses loans to a group prevents its members from building the credit history that would change the model's view. A recommendation system that predicts you'll prefer a certain kind of content shows you only that, making the prediction true. The philosopher Ian Hacking called this the looping effect of human kinds: unlike stars or atoms, people react to how they are classified, and the classifications change as a result.

§ 5Work, automation and hidden labour

How many jobs?

Forecasts of automation have a long and mixed record. A widely cited 2013 paper by Carl Benedikt Frey and Michael Osborne estimated that 47% of US employment was at high risk of computerisation. A 2016 OECD study that looked at the tasks within occupations, rather than whole occupations, put the share of jobs at high risk at around 9% across rich countries. The difference was method: jobs are bundles of tasks, and automating some tasks usually changes a job rather than eliminating it.

Generative AI changed which tasks were exposed. Earlier automation mainly threatened routine manual and clerical work. A 2023 analysis by Tyna Eloundou and colleagues estimated that around 80% of the US workforce had at least 10% of their tasks exposed to large language models, and around 19% had at least half — with exposure higher for better-paid, more educated occupations. An International Labour Organization study the same year (Gmyrek, Berg and Bescond) found that the most exposed category globally was clerical work, that the main effect was more likely to be augmentation than full automation, and that women were more exposed than men because they are over-represented in clerical jobs, especially in higher-income countries. The International Monetary Fund estimated in January 2024 that almost 40% of global employment was exposed to AI, rising to around 60% in advanced economies.

Exposure is not displacement. Whether exposed tasks lead to lost jobs, changed jobs or new jobs depends on choices by firms, workers, unions and governments — which is to say, on social and political arrangements. Early experimental studies, covered in the studies section, found productivity gains concentrated among less experienced workers. What those gains mean for wages and employment is not yet known.

Deskilling, upskilling, and the moral crumple zone

Harry Braverman's Labor and Monopoly Capital (1974) argued that management uses technology to strip skill and control from workers. Shoshana Zuboff's 1988 study offered a more open view: technology could “automate” or “informate”. Both possibilities are visible with AI. A system that captures the know-how of the best customer service agents and gives it to novices can be read as upskilling the novices — or as extracting the tacit knowledge of experienced workers and making them replaceable.

The anthropologist Madeleine Clare Elish introduced a useful concept for what happens when humans and automated systems share control: the moral crumple zone (2019). Just as a car's crumple zone absorbs the impact of a crash, the human operator in a highly automated system absorbs the moral and legal responsibility when things go wrong — even when they had little real control. The “human in the loop” that regulations require can become a person whose job is to take the blame.

Algorithmic management

For millions of workers, the boss is already partly software. Alex Rosenblat's Uberland (2018) showed how ride-hail drivers are managed through ratings, nudges, opaque pay algorithms and the threat of “deactivation”. Katherine Kellogg, Melissa Valentine and Angèle Christin's review “Algorithms at Work” (2020) identified six mechanisms of algorithmic control — restricting, recommending, recording, rating, replacing and rewarding — and documented the ways workers resist: gaming metrics, collective information-sharing, and refusal. These are Goffman's “secondary adjustments”, and the Hawthorne workers' informal output norms, in a new setting.

Ghost work

“Artificial” intelligence depends on enormous amounts of human labour. Mary Gray and Siddharth Suri's Ghost Work (2019) documented the on-demand workforce of people who label data, transcribe, moderate content and correct machine outputs through online platforms, typically paid per task and invisible to end users. Sarah T. Roberts's Behind the Screen (2019) studied commercial content moderators, including in Manila, who spend their days viewing violent and abusive material so that others don't have to see it.

Investigations have shown how this works for generative AI. In January 2023, TIME reported that workers in Kenya employed through an outsourcing firm had been paid less than US$2 an hour to label graphic descriptions of violence and abuse, so that a chatbot could learn to avoid producing such content. In 2023, The Washington Post reported on Filipino workers doing annotation through the platform Remotasks, many of whom described low pay, payment delays and withheld wages. As models have moved towards expert-level tasks, a newer layer of better-paid annotation work by graduates, doctors and lawyers has emerged alongside the old one — but the basic structure, a distributed, contingent workforce invisible to users, remains.

What Marienthal tells us

Suppose AI does replace a large share of current work, as some in the industry predict. The common answer is a universal basic income. The Marienthal study (see the studies) and Jahoda's later research suggest that income is only one of the things work provides. Time structure, social contact beyond the family, participation in collective purposes, status and identity, and regular activity would all need to come from somewhere else. That is a sociological design problem, not an economic one, and almost no one is working on it.

§ 6Data colonialism and the Global South

Nick Couldry and Ulises Mejias's The Costs of Connection (2019) argue that the large-scale extraction of personal data is best understood as a new phase of colonialism: not a metaphor, they insist, but a continuation of colonialism's basic move, the appropriation of resources by declaring them free for the taking. Where historical colonialism appropriated land and bodies, data colonialism appropriates human life itself, as data. Critics argue the analogy minimises the violence of historical colonialism. But the structural pattern it describes, with raw material and cheap labour flowing from the periphery and finished products and profits concentrating in the core, closely follows the world-systems map (see Theory and politics).

Language is where this is most visible. Most of the text used to train major language models is in English, followed by a handful of other widely written languages. Hundreds of languages spoken by millions of people — including most of the more than one hundred languages of the Philippines — are thinly represented. Models consequently work worse in those languages, and encode the cultural assumptions of the dominant ones. A 2023 study by Mohammad Atari and colleagues asked “Which humans?” and found that large language model responses to psychological survey items most resembled those of people in Western, educated, industrialised, rich and democratic (WEIRD) societies.

Responses include community efforts such as Masakhane, a grassroots network building natural language processing for African languages; regional projects such as AI Singapore's SEA-LION models for Southeast Asian languages; and government strategies framed around digital sovereignty. The open question, in Alatas's terms, is whether such efforts will produce models that think differently, or simply translations of a captive mind.

§ 7Knowledge, truth and the public sphere

The social construction of machine knowledge

Berger and Luckmann argued that what a society takes as reality is built through interaction, institutionalised, and then experienced as objective. Large language models are a new kind of participant in this process. They are trained on what a society has written, they answer questions in a confident, institutional voice, and their answers then flow back into what society writes. Millions of people now ask a chatbot before they ask a person, a library or a search engine, which makes the companies that train these models a new kind of knowledge institution, with the power to frame what seems like common sense — Gramsci's hegemony, in a technical form.

There is evidence for one predictable effect: homogenisation. A 2024 study in Science Advances by Anil Doshi and Oliver Hauser found that short stories written with AI assistance were rated as more creative individually, but were more similar to each other than stories written without it. Individual gain, collective loss of diversity: a classic social dilemma.

There is also a risk of feedback. A 2024 study in Nature by Ilia Shumailov and colleagues showed that models trained recursively on text generated by earlier models progressively lose information about the rarer parts of the original distribution — “model collapse”. As more of the web is generated by machines, what happens to the tails of human experience in future training data?

Persuasion, synthetic media and the liar's dividend

Fears that AI would flood elections with persuasive disinformation have so far run ahead of the evidence. Research on elections in 2024 found many examples of AI-generated content but limited evidence that it changed outcomes, consistent with the long sociological finding that direct media persuasion is limited (see The People's Choice). But the same research tradition says influence works through trust, networks and repetition rather than single messages, and AI makes producing these at scale very cheap.

Two findings point in different directions. A 2024 study in Science by Thomas Costello, Gordon Pennycook and David Rand found that a few rounds of conversation with GPT-4, tailored to the participant's own stated evidence, reduced belief in conspiracy theories by about 20% on average, with effects lasting at least two months. Conversational AI can persuade, sometimes towards accuracy. The same capacity could be used to persuade towards anything.

The legal scholars Bobby Chesney and Danielle Citron identified a subtler risk in 2019: the liar's dividend. When everyone knows convincing fakes are possible, people caught on genuine recordings can claim they are fake. The damage of synthetic media may be less that people believe false things than that they stop believing true ones — an erosion of the shared evidential ground on which, Habermas would say, a public sphere depends.

§ 8Intimacy, care and loneliness

In 1966, the MIT computer scientist Joseph Weizenbaum built ELIZA, a simple program that mimicked a psychotherapist by reflecting users' statements back as questions. He was disturbed to find that people, including his own secretary, confided in it and wanted privacy with it. His Computer Power and Human Reason (1976) argued that there are tasks computers should not be given, whether or not they can do them, because they require human judgement and care. The tendency to attribute understanding to systems that mimic conversation is still called the ELIZA effect.

Sherry Turkle's Alone Together (2011) extended the argument to social robots and networked life: technologies that offer “the illusion of companionship without the demands of friendship”. AI companion apps, which offer an always-available, endlessly patient friend or partner, now have many millions of users. When the companion app Replika abruptly restricted romantic and sexual role-play in early 2023, many users described grief akin to bereavement — a vivid sign of how real these attachments had become.

The evidence on effects is early and mixed. Some studies find that conversations with AI companions reduce feelings of loneliness in the short term. Research published in 2025 by MIT Media Lab and OpenAI found that heavier daily use of a chatbot was associated with greater loneliness and emotional dependence, though the direction of causation is unclear: lonely people may use chatbots more. Concerns about children and vulnerable users have led to lawsuits in the US and to proposals for age limits and safeguards in several countries.

Sociology offers three questions here. From Hochschild: what happens when care becomes a product designed to maximise engagement, performing feeling without having any? From Durkheim: does a companion that always agrees provide the friction — the obligations, disagreements and reciprocity — that integrate people into society, or substitute for it? From the sociology of care: who will look after the elderly and isolated, and will AI companions supplement human care or become the excuse for withdrawing it?

§ 9Surveillance and power

Foucault's panopticon is the most over-used image in the study of digital technology, and some of its critics point out why: the panopticon disciplined people by making them aware of being watched, whereas digital surveillance often works precisely because people forget it. The sociologist David Lyon prefers the idea of surveillance culture, in which people are not only watched but also watch, share and participate willingly.

Shoshana Zuboff's The Age of Surveillance Capitalism (2019) offered the most influential economic account: firms extract “behavioural surplus” — data beyond what is needed to provide a service — and use it to predict and modify behaviour for sale to advertisers. Critics have questioned whether targeted advertising is as effective at modifying behaviour as Zuboff suggests, but her account of the business logic is widely accepted.

State surveillance with AI is advancing quickly, particularly with facial recognition. China's “social credit system” is often described in Western media as a single, all-seeing score for every citizen. Researchers who have studied it closely describe something more fragmented: a patchwork of local pilot schemes, corporate credit ratings and court blacklists, focused heavily on businesses and on debt enforcement. That is not reassurance — China's surveillance of Uyghurs in Xinjiang, involving extensive use of biometric and data systems, has been extensively documented — but it is a reminder that a sociologically accurate account of surveillance requires actual study, not a metaphor.

§ 10Governance: who sets the rules

Rules for AI are being written now, and their shape reflects the political traditions described in Theory and politics. The details below are current as of October 2026; this area changes fast.

Three sociological observations about this landscape:

1. Regulation is being written in the language of risk. That reflects Beck's risk society and a long European tradition. Risk framing tends to hand authority to technical experts and standards bodies, and to treat harms as probabilities to be managed rather than rights to be protected. The EU's delay of its high-risk rules was justified in part because the technical standards and national authorities were not ready — an example of how much the real content of law is decided in the less visible world of standards-setting.
2. Audit is becoming the default remedy. The accounting scholar Michael Power warned in The Audit Society (1997) that audits can become rituals of verification that produce comfort rather than accountability. Algorithmic audits are valuable, as Gender Shades showed, but who audits, with what access, and with what power to enforce, decides whether they are more than ritual.
3. The real governance is often private. Usage policies, model specifications, content rules and the choices of a few companies about what their systems will and will not do govern far more interactions every day than any statute. This is what elite theorists and Gramscians would predict, and what deliberative democrats would want to open up.

§ 11AI as a research instrument

AI is also changing how sociology is done. Since the 2009 manifesto for computational social science, researchers have used machine learning to analyse text, images and networks at scales no team of human coders could manage. Large language models have extended this: they can classify open-ended survey answers, code interview transcripts, and extract information from historical documents, often with accuracy comparable to trained human coders, at a fraction of the cost.

A more controversial idea is to use language models as synthetic respondents. In 2023, Lisa Argyle and colleagues showed in Political Analysis that a language model, given demographic backstories, could reproduce many patterns in American public opinion survey data. They called the result a “silicon sample”. Subsequent work has been more sceptical. Studies found that such samples reproduce averages better than variation, are less accurate for minority groups and for non-Western populations, and can change answers unpredictably as models are updated, which threatens the reproducibility on which science depends. Researchers have also built simulations populated by language-model “agents” to explore social dynamics, an extension of the agent-based modelling tradition, with the same question hanging over them: is this a model of society, or a model of the internet's description of society?

§ 12Open questions

The questions below are ones no one has satisfactory answers to yet. They are where the sociology of AI is heading, and many of them could be studied with the tools described elsewhere on this site.

1. Distribution. Who captures the productivity gains from generative AI: firms, high-skill workers, low-skill workers, consumers? How does this differ between countries with strong and weak labour institutions?
2. The content of work. When AI does the routine part of a job, is the remainder more meaningful, or more intense and more closely monitored?
3. Expertise and authority. What happens to professions — medicine, law, teaching, journalism — whose claim to authority rests on specialised knowledge that machines can now partially reproduce?
4. Socialisation. How does growing up with conversational AI change how children learn language, form relationships and develop a sense of what is true?
5. Norms. How are new norms around authorship, honesty and disclosure of AI use forming, and who is enforcing them? Durkheim would ask: where is the anomie, and what will fill it?
6. Inequality of access and of exposure. Who gets AI as a tool that serves them, and who gets AI as a system that judges them?
7. Collective action. Can workers, communities and countries organise to shape AI, as the Writers Guild did? What does solidarity look like among a dispersed, platform-mediated workforce?
8. Knowledge. If more and more of what people read is generated or filtered by a few models, what happens to the diversity of ideas a society has available to it?
9. Legitimacy. Under what conditions do people accept decisions made or shaped by machines as legitimate? Weber's question, newly urgent.

The most important thing sociology can bring to AI is the refusal to treat its future as already decided. Technologies have politics, but they do not have destinies.