• Noble, Safiya Umoja. Algorithms of Oppression: How Search Engines Reinforce Racism. New York University Press, 2018.
• Benjamin, Ruha. Race After Technology: Abolitionist Tools for the New Jim Code. Polity, 2019.
• Buolamwini, Joy, and Timnit Gebru. “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” FAT 2018.
• D’Ignazio, Catherine, and Lauren F. Klein. Data Feminism. MIT Press, 2020.
• Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.
• Gray, Mary L., and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019.
• Roberts, Sarah T. Behind the Screen: Content Moderation in the Shadows of Social Media. Yale University Press, 2019.
• Selwyn, Neil. Should Robots Replace Teachers? AI and the Future of Education. Polity, 2019.
• Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin’s Press, 2018.
• O’Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown, 2016.
• Herzfeld, Noreen L. The Artifice of Intelligence: Divine and Human Creativity in the Age of Thinking Machines. Fortress Press, 2023.
• Reed, Esther D. Digital Souls: A Theology of Artificial Intelligence. SCM Press, 2021.
• Pasquale, Frank. The Black Box Society: The Secret Algorithms That Control Money and Information. Harvard University Press, 2015.
• Zuboff, Shoshana. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs, 2019.
• Online resources: Princeton SOC 426 materials; MIT STS.084 syllabus; NYU AI Now Institute annual reports; Edinburgh Data Society programme; Oxford Internet Institute working papers; Data & Society Research Institute publications.

Chapter 1: Sociology Meets AI — Why Social Science Matters

The encounter between sociology and artificial intelligence is not, at its core, a story about machines becoming more human. It is a story about humans — the specific humans who build AI systems, the specific institutional contexts in which those systems are developed and deployed, and the specific populations who bear the benefits and costs of their operation. Science and Technology Studies (科学技术研究, or STS) emerged in the 1970s and 1980s as a discipline dedicated to precisely this kind of inquiry: the social shaping of technical artefacts and systems. Where earlier frameworks tended toward technological determinism (技术决定论) — the view that technology has inherent properties that drive social change in predictable directions — STS scholars argued that the outcomes of technical systems are always contingent on social choices made at every stage of design, deployment, and use. The same technology can produce radically different social effects depending on who controls it, who is subject to it, and what institutional scaffolding surrounds it.

Artificial intelligence presents a particularly rich and urgent case for this kind of analysis. The term itself is contested and historically variable: what counted as “artificial intelligence” in 1956, when the Dartmouth conference coined the phrase, differs substantially from what is called AI in 2025, when the term most often refers to large-scale machine learning systems trained on vast datasets. But across these definitional shifts, one sociological constant holds: AI systems are built by people, trained on data that people produced, deployed in institutional contexts that people designed, and applied to populations whose social characteristics shape every aspect of the system’s behaviour and impact. The claim that AI systems are neutral tools that simply “crunch data” and produce objective results is, from a sociological perspective, not merely incomplete but actively misleading. It obscures the choices — about what data to collect, how to label it, what outcome to optimise for, whose interests to prioritise — that are embedded in every system and that have profound distributional consequences.

Frank Pasquale’s metaphor of the black box (黑盒) is a useful entry point. In Pasquale’s account, the algorithms that control credit, employment, insurance, and information are opaque by design: their inputs, weights, and decision logic are proprietary trade secrets, inaccessible to the people they affect and often to the regulators nominally charged with overseeing them. The sociological project of this course is in part an exercise in opening the box — not to find clean mathematical operations inside, but to find the social world. Inside the black box of a predictive policing algorithm, a sociologist finds training data drawn from historically biased arrest records, outcome variables defined by police departments with specific institutional interests, and deployment decisions made by city governments under fiscal pressure. Inside the black box of a resume-screening system, a sociologist finds labelled examples of “successful” employees drawn from firms whose historical hiring practices excluded women and people of colour. The mathematics is real, but it is mathematics applied to social material by social actors for social purposes.

The course proceeds through a series of substantive domains: algorithmic inequality and the replication of disadvantage across welfare, housing, and criminal justice; race and gender bias in facial recognition and search; AI’s transformation of the media and information environment; the AI-powered classroom and its limits; the hidden human labour that animates apparently automated systems; the theological and religious dimensions of AI’s cultural reception; surveillance capitalism and the governance of algorithmic power; and the emerging frameworks of intersectional and design-justice critique. Each domain illuminates a different facet of AI as a sociological phenomenon, and each draws on a distinct body of empirical and theoretical scholarship. Across all of them, the underlying question is the same: who benefits, who is harmed, and by what mechanisms does AI reshape the social world?

The course is in conversation with several adjacent disciplines that approach these questions from different angles. HIST 415 (History of AI) traces the historical genealogy of today’s systems, showing how contemporary AI’s racial and gender pathologies are not accidents but continuities with longer histories of classification, enumeration, and administrative power. PHIL 451 (AI Ethics, Law, and Governance) works at the normative and regulatory frontier, asking what legal and ethical frameworks are adequate to AI’s challenges. The distinctly sociological contribution is structural: the insistence that AI cannot be understood as a series of individual decisions by engineers and executives but must be analysed as a phenomenon embedded in, and generative of, patterns of inequality, power, and social reproduction that exceed any individual actor’s intentions or awareness. This structural analysis is what the course’s twelve weeks are dedicated to developing.

Chapter 2: Algorithmic Inequality — How AI Systems Produce and Reproduce Disadvantage

Cathy O’Neil’s Weapons of Math Destruction provides the most widely cited diagnostic framework for understanding how algorithmic systems generate harm at scale. O’Neil identifies three features that characterise what she calls a “WMD”: the system is opaque, meaning those it affects cannot understand or contest how it works; it operates at scale, meaning its effects are not localised but spread across entire populations; and it causes damage, particularly to the people it ranks lowest. The three features interact: scale amplifies harm, opacity prevents correction, and the combination creates systems that are simultaneously very powerful and very difficult to hold accountable. O’Neil’s case studies range across credit scoring systems that treat zip code as a proxy for creditworthiness (which, given residential segregation, functions as a proxy for race), to teacher evaluation systems that fire effective educators on the basis of statistically unreliable value-added metrics, to the recidivism prediction tools used by courts in sentencing decisions. In each case, the algorithm claims objectivity — it is just numbers, just data — while encoding the historical inequalities that the data reflects.

Virginia Eubanks extends this analysis with specific attention to how automated decision systems are applied to low-income populations in the United States. Her concept of the digital poorhouse (数字济贫院) draws on the history of nineteenth-century poorhouses — institutions that concentrated surveillance, discipline, and punishment on the poor under the guise of welfare — to characterise the function of contemporary automated systems. The Indiana automated welfare eligibility system terminated benefits for hundreds of thousands of low-income residents on the basis of algorithmic determinations of non-compliance, with errors that disproportionately affected people with disabilities, non-English speakers, and people experiencing housing instability. The Allegheny County child welfare algorithm scores families for risk of child abuse, with training data derived from families who had previous contact with child protective services — meaning the system is measuring not the underlying probability of harm but the probability of prior surveillance. These are not mere technical errors; they are the systematic application of algorithmic power to populations who were already subject to heightened administrative scrutiny before any algorithm was involved.

A key dynamic in understanding how algorithmic systems reproduce inequality is the feedback loop (反馈回路). Predictive policing offers the clearest illustration. A system like PredPol trains on historical arrest data to predict where crime is likely to occur, directing patrol resources to those locations. Officers dispatched to predicted crime areas make arrests, which are recorded in the data, which reinforce the prediction that crime is high in those areas, which directs more officers there, which produces more arrests. The system does not measure crime; it measures policing. If crime were distributed uniformly across a city but policing were historically concentrated in low-income neighbourhoods of colour, a system trained on arrest data would perpetuate exactly that geographic and demographic concentration, generating the appearance of an objective, data-driven confirmation of a pattern that was actually produced by the historical decisions of police departments. Bernard Harcourt, among others, has demonstrated this dynamic formally, but Eubanks and O’Neil make the political stakes vivid through concrete case studies.

The distributional justice (分配正义) problem in algorithmic systems cuts deeper than accuracy. A system that accurately predicts which loan applicants are likely to default based on historical repayment data may still be unjust if low-income applicants’ poor repayment histories are themselves the product of structural disadvantage — lower wages, more volatile employment, higher housing costs relative to income — rather than any intrinsic tendency toward non-repayment. The accurate prediction reproduces and reinforces the disadvantage it reflects. This is the sense in which O’Neil argues that WMDs are not merely inaccurate systems in need of technical improvement, but systems that encode a specific theory of social causation — that past behaviour predicts future behaviour, full stop — while ignoring the structural conditions that produced that behaviour and that the system, by denying credit, housing, or employment, will perpetuate.

The legal distinction between disparate impact (不平等影响) and disparate treatment (不平等对待) is important here. Disparate treatment refers to intentional discrimination — a lender who refuses loans to Black applicants because of their race. Disparate impact refers to facially neutral policies that produce racially unequal outcomes — a lender whose credit scoring model, applied without racial intent, denies loans to Black applicants at much higher rates. US anti-discrimination law, developed largely in the employment context, has historically been more effective at addressing disparate treatment than disparate impact, and algorithmic systems primarily generate the latter. An algorithm does not intend to discriminate; it optimises for an outcome using features that correlate with protected characteristics without explicitly using those characteristics, a process that the legal scholar Solon Barocas and Moritz Hardt have called “redundant encoding.” The gap between the legal tools available and the form of discrimination that algorithmic systems actually produce is a central governance challenge, and one that the EU AI Act and the New York City automated employment decisions law represent early attempts to address, with significant limitations that Chapter 8 examines in detail.

Chapter 3: Race, Gender, and Algorithmic Bias

The Gender Shades study conducted by Joy Buolamwini and Timnit Gebru in 2018 is one of the most consequential empirical contributions to the sociology of AI. Buolamwini and Gebru evaluated three commercial facial analysis systems — products of Microsoft, IBM, and Face++ — for accuracy in classifying the gender of faces. The systems performed well overall, but their performance varied dramatically across demographic categories: while lighter-skinned men were classified correctly more than 99 percent of the time, darker-skinned women were misclassified at rates as high as 35 percentage points worse. The disparity was not marginal; it was enormous. And the systems had been deployed commercially with no disclosed evaluation of their accuracy across demographic groups. The study prompted Microsoft and IBM to acknowledge the problem and release improved systems, but it also revealed the culture of AI development — a culture in which accuracy on benchmark datasets was the measure of success, and the demographic composition of those datasets was not systematically interrogated.

Safiya Umoja Noble’s Algorithms of Oppression examines a different but related phenomenon: the politics of search ranking. Noble’s starting point is an empirical observation — that in 2011, typing “Black girls” into Google returned pornographic results in the top positions, while typing “white girls” returned more neutral content. The disparity was not a quirk of Noble’s individual search history; it was a systematic feature of Google’s ranking algorithm, which optimised for click-through rates and advertising revenue rather than for the accuracy, dignity, or representational fairness of results. Noble’s analysis extends beyond this example to examine how search algorithms systematically amplify racist, sexist, and otherwise demeaning representations of marginalised communities, because those representations generate engagement that the algorithm’s commercial incentives reward. The key theoretical move is to treat search not as a neutral window onto information but as a publisher with a politics — a politics that is commercial, racialised, and gendered.

Ruha Benjamin’s Race After Technology synthesises a wide range of empirical cases into a theoretical framework she calls the New Jim Code (新吉姆代码). The name is deliberate and polemical: Jim Crow laws were the formally codified system of racial hierarchy in the post-Reconstruction American South, and Benjamin argues that ostensibly neutral technical systems perform analogous work today, encoding racial hierarchy in a form that is even harder to contest because it presents itself as objective mathematics rather than explicit racial categorisation. The New Jim Code encompasses facial recognition systems that fail for darker-skinned faces, predictive policing systems that concentrate surveillance on communities of colour, and health algorithms that systematically underestimate the health needs of Black patients — as in the Optum insurance algorithm that used health expenditures as a proxy for health needs, systematically underestimating the needs of Black patients who, because of structural barriers to healthcare access, spent less on healthcare despite equivalent or greater health needs. The connection to Jim Crow is not just rhetorical: Benjamin traces how the specific technical systems of today are connected, through institutional histories and personnel, to earlier forms of racially discriminatory administration.

Catherine D’Ignazio and Lauren Klein’s Data Feminism brings an explicitly feminist framework to bear on these questions, drawing on Patricia Hill Collins’s concept of the matrix of domination (支配矩阵) — the interlocking systems of race, gender, class, and other axes of power that organise social life and that feminist analysis must address simultaneously rather than in isolation. D’Ignazio and Klein apply this framework to data science and AI, asking: who has the power to collect data? On whom? For what purposes? Whose experiences and perspectives are reflected in the categories that data systems use? The answers consistently reveal a pattern: data science overwhelmingly reflects the perspectives and interests of those with social power — predominantly white, predominantly male, predominantly wealthy, predominantly Western — while treating those perspectives as universal and objective. The feminist data science programme they articulate is not simply about adding diversity to existing pipelines; it is about transforming the epistemological foundations of data practice to recognise whose knowledge counts and whose experiences shape what gets measured.

The common framing of these problems as “bias” is worth interrogating carefully. The term implies a deviation from a neutral baseline — a biased scale is one that diverges from the true weight of an object. But there is no neutral AI system waiting to be recovered by removing bias. The training data reflects the world that has been built by historically specific practices of photography, surveillance, categorisation, and data collection — a world in which lighter-skinned individuals have been disproportionately photographed under conditions that train facial recognition systems well, and in which the digitisation of records has been faster and more complete for wealthy, Western populations. This is not a deviation from neutrality; it is the baseline. The question is not how to restore neutrality but how to negotiate the inescapable fact that all AI systems encode some values, some epistemological commitments, some vision of who counts and what matters — and to ask whether those values are ones that can be openly defended and democratically chosen.

Chapter 4: AI, Media, and the Information Environment

Shoshana Zuboff’s The Age of Surveillance Capitalism provides the foundational theoretical framework for understanding how AI systems have transformed the economics and politics of media. Zuboff’s central claim is that a new economic logic — surveillance capitalism (监控资本主义) — emerged in Silicon Valley in the early 2000s and has since become dominant across the digital economy. Surveillance capitalism rests on the extraction of what Zuboff calls “behavioural surplus” from human activity: the data generated by online behaviour that exceeds what is needed to improve the immediate service and is instead processed by machine learning systems to generate predictive products — predictions about what individuals will do, buy, click, or vote for. These predictions are sold to advertisers seeking to influence behaviour, creating a fundamentally new kind of market in human futures. The apparatus of prediction — the data centres, the machine learning models, the feedback loops from ad clicks to model updates — is AI infrastructure in the most literal sense, and its economic logic shapes the media environment in which billions of people now live.

The filter bubble (过滤气泡) hypothesis, popularised by Eli Pariser in 2011, proposes that algorithmic curation — the use of machine learning systems to select which content each user sees on social media, news aggregators, and search engines — systematically isolates users in ideologically homogeneous information environments, narrowing exposure to diverse perspectives and contributing to political polarisation. The empirical status of this claim is contested. Researchers including Seth Flaxman and Levi Boxell have found evidence of personalisation effects, while others like Axel Bruns and Brundidge et al. have found that social media users are actually exposed to more diverse political content than those who rely on traditional media. The discrepancy may partly reflect methodological differences — what counts as exposure, how diverse perspectives are measured — but it also reflects the genuine complexity of how recommendation algorithms interact with user behaviour, social networks, and the content that publishers choose to produce. What is less contested is that algorithmic recommendation systems are optimised for engagement, and that outrage, fear, and moral condemnation are reliably more engaging than nuanced analysis — an incentive structure with implications for the character of political discourse regardless of whether strict filter bubbles form.

Generative AI has introduced a qualitatively new dimension to the information environment: the capacity to produce synthetic text, images, audio, and video that are increasingly indistinguishable from human-produced content. The implications for disinformation are substantial. Deepfakes (深度伪造) — AI-generated synthetic video of real individuals appearing to say or do things they did not — have been used in political disinformation campaigns and in non-consensual intimate image abuse. Synthetic text generated by large language models can produce convincing misinformation at scale, overwhelming fact-checking capacity. But the political scientist Bobby Chesney and Danielle Citron have identified what they call the “Liar’s Dividend” — the possibility that the existence of convincing deepfakes allows political actors to dismiss authentic video evidence as fabricated, eroding the evidentiary value of documentation regardless of whether any specific document has been manipulated. The social epistemology problem posed by AI-generated synthetic media is thus not only about specific instances of fabrication but about the broader destabilisation of shared epistemic foundations.

AI’s relationship to journalism is multifaceted and does not reduce to simple displacement or augmentation. News organisations including the Associated Press and Reuters have used automated writing systems to generate structured stories — earnings reports, sports results, weather updates — for more than a decade. These systems are not general-purpose language models but template-filling systems that convert structured data into prose following a fixed format; they free journalists for work requiring judgment and contextual knowledge. More recently, general-purpose language models have entered newsrooms as productivity tools for drafting, summarising, and transcription. But AI also shapes journalism through the economics of digital media: algorithmic distribution by social media platforms determines what journalism reaches audiences, creating incentives for publications to produce content that platforms will recommend — a dynamic that has contributed to the hollowing out of local news, the proliferation of outrage-optimised content, and the concentration of digital advertising revenue in the platforms rather than the publications that produce the journalism.

Platform power over content governance is the point at which apparently automated AI systems most clearly reveal their human dimensions. Sarah Roberts’s research on content moderation (内容审核) demonstrates that the moderation of social media content — decisions about what speech to allow and what to remove — is performed not only by automated classifiers but by a large and largely invisible human workforce. These workers, employed by outsourcing firms in the Philippines, India, Kenya, and elsewhere, review flagged content that automated systems cannot confidently classify, making decisions under severe time pressure, with exposure to the most disturbing material on the platform, and with limited institutional support for the psychological consequences of that exposure. The moderation workforce reveals something important about the limits of AI in high-stakes classification tasks and about the distribution of cognitive labour in the global economy — themes that Chapter 6 develops at length in the context of data labelling and platform labour more broadly.

Chapter 5: AI and Education — Reshaping Teaching and Learning

Neil Selwyn’s Should Robots Replace Teachers? is notable for its careful empirical temperament in a domain prone to both utopian and dystopian excess. Selwyn does not argue that AI in education is simply good or simply bad; he argues that the specific claims made by AI education vendors — that their systems personalise learning, improve outcomes, increase engagement, and reduce inequality — are not adequately supported by evidence, and that the enthusiasm with which these claims have been accepted by policymakers and education administrators reflects the ideological attraction of technocratic solutions to structural educational problems more than rigorous evaluation of what the systems actually do. This is a distinctly sociological argument: the adoption of AI in education is not driven purely by evidence of effectiveness but by institutional dynamics, market pressures, and ideological commitments that require sociological analysis.

Adaptive learning systems (自适应学习系统) represent the most substantive AI application in formal education. Systems like Knewton, DreamBox, and Carnegie Learning’s MATHia track students’ responses to practice problems and adjust the difficulty and type of content presented based on inferred models of each student’s knowledge state. The underlying technology — Bayesian knowledge tracing, item response theory, reinforcement-learning-based curriculum sequencing — is sophisticated, and the systems have generated genuine improvements in measurable learning outcomes in specific domains, particularly mathematics at the K-12 level. But Selwyn and other critical education technology scholars identify important limitations. Adaptive systems measure what is measurable — response time, answer correctness, click sequences — and optimise for what they can measure. The aspects of learning most valued by educators — curiosity, critical thinking, collaborative problem-solving, the construction of personal meaning from content — are precisely the aspects that are hardest to measure and therefore most likely to be neglected or crowded out by the optimisation pressure that adaptive systems create.

The transformation of assessment through AI raises particularly acute questions. Automated essay scoring systems, in use since the 1990s but increasingly sophisticated, evaluate student writing using features that correlate with human scores — syntactic complexity, vocabulary range, essay length, semantic coherence — without understanding the content of the essay. Research has shown that these systems can be “gamed” by students who produce structurally complex but substantively incoherent text, and that they perform less well on creative, unconventional, or culturally specific writing. The emergence of capable large language models capable of producing fluent, contextually appropriate prose has supercharged the assessment crisis: students can now submit AI-generated text that is, in many respects, better than what they would have written themselves, while detection tools remain unreliable and frequently misidentify human writing as AI-generated. Universities and professional examination bodies are responding with a mixture of policy statements, detection tools, and pedagogical redesign, but no consensus has emerged on what assessment should look like in a world of capable text generators.

Selwyn’s most important argument concerns what he calls the pedagogical relationship (教育关系). Education is not, at its core, an information-transfer problem that can be optimised by a sufficiently capable personalisation system. It is a relational and developmental process: students learn who they are and what they can do partly through their interactions with teachers, who model intellectual engagement, maintain high expectations, provide emotionally calibrated feedback, and participate with students in the construction of shared meaning within a community of practice. An AI tutor that optimises for correct answers on measurable assessments may, in doing so, undermine the relational and developmental dimensions of education — dimensions that are not captured in the outcome metrics the system is optimised for, but that are arguably more important in the long run. This is not an argument against AI in education; it is an argument for a more comprehensive account of what education is for before deciding what role AI should play in it.

The equity dimensions of AI in education deserve specific attention. Adaptive learning systems require data to personalise effectively, and more data enables better personalisation — a dynamic that advantages students in better-resourced schools, who have been using the system longer and whose performance data is more complete. This is the opposite of the equity improvement that vendors often claim. Moreover, the deployment of AI tutoring and automated assessment tends to concentrate human teacher attention at the top of the distribution — students who are already performing well benefit from the resources freed by AI handling routine practice and assessment — while struggling students may receive more AI-mediated instruction and less human contact precisely when human relationship and support are most important. The equity effects of AI in education are thus not uniformly positive or negative but are mediated by institutional choices about how AI tools are deployed, choices that are themselves shaped by the resource constraints and institutional priorities that track existing inequalities.

Higher education is experiencing a governance crisis over AI that is particularly visible in the domain of written coursework. University policies have ranged from blanket prohibitions on AI tool use (largely unenforceable) to unrestricted permission to use whatever tools are available (which raises questions about what is being assessed) to nuanced frameworks that distinguish between AI use for idea generation, drafting, and editing. The underlying question — what is a university education for in a world where capable AI can perform many of the tasks previously reserved for educated humans — is not a question that can be answered by a policy committee. It requires a broader sociological and philosophical analysis of the purposes of higher education, the labour market signals that credentials send, and the kinds of cognitive and social capacities that universities are, or should be, in the business of developing.

Chapter 6: The Hidden Workforce — Ghost Work, Content Moderation, and Data Labour

Mary Gray and Siddharth Suri’s Ghost Work takes its title from the anthropologist’s concept of ghost labour — work that is structurally necessary to a social system but systematically rendered invisible by that system’s self-representation. In the context of the contemporary AI economy, ghost work refers to the millions of human micro-tasks that are abstracted behind application programming interfaces (APIs) and presented to consumers as automated services. When a user speaks to a voice assistant and the response is highly accurate, part of that accuracy reflects human workers who have labelled training data, corrected transcription errors, and evaluated response quality. When a mapping application identifies a business from a street photograph, human workers have annotated thousands of similar photographs to train the recognition model. When a platform recommends content that matches a user’s preferences, human workers have rated the relevance and quality of millions of content items. The API economy (API经济) makes all of this human labour invisible by design: the API presents a clean interface that abstracts away the infrastructure behind it, and the infrastructure is presented as AI to conceal its human component.

The geography of ghost work is not accidental. The major platforms — Amazon Mechanical Turk, Microworkers, Scale AI, Surge AI — draw their labour supply from countries where the wage gap with the United States is large enough that micro-tasks paying a few cents are economically viable. Workers in the Philippines, India, Kenya, Bangladesh, and Venezuela perform significant proportions of AI data labour for wages that reflect not the value they add to billion-dollar AI systems but the economic legacy of colonial extraction that has produced wage differentials across the global economy. Gray and Suri document the working conditions of these workers in detail: the algorithmic management systems that review work quality and deactivate workers for errors without explanation or appeal; the constant uncertainty of task availability that prevents income planning; the absence of labour protections that apply to employees; and the social isolation of piece-rate work distributed across thousands of independent workers who have little capacity for collective action. The invisibilisation of this labour is not only an economic fact; it is ideological — the “AI” framing actively serves corporate interests by making the human workforce beneath the platform harder to see, organise, and regulate.

Sarah Roberts’s Behind the Screen provides one of the most detailed accounts of a specific form of AI labour: commercial content moderation (商业内容审核). Content moderators review material flagged by users or automated classifiers that may violate platform policies — graphic violence, child sexual abuse material, terrorist propaganda, hate speech — and make decisions about whether to remove it. Roberts conducted ethnographic fieldwork with moderators in the United States, the Philippines, and Silicon Valley, and her account reveals the human costs of this work in stark terms. Moderators develop what Roberts calls “secondary traumatic stress” — the psychological symptoms associated with trauma exposure in the absence of direct personal experience — from sustained exposure to the most disturbing material that billions of people upload to social media platforms. The platforms have been slow to acknowledge this harm, to provide adequate mental health support, or to limit exposure volumes, in part because doing so would require acknowledging the existence of the human workforce that platform presentations of “AI moderation” are designed to obscure.

The resistance of AI data workers has taken several forms. In 2019, content moderators employed by Cognizant on behalf of Facebook filed a class action lawsuit alleging inadequate mental health support and exposure to traumatic content. In Kenya, workers employed by Sama to label data for OpenAI organised and ultimately brought a lawsuit alleging that they were paid poverty wages to label graphic descriptions of sexual violence and terrorism. The European Union’s Platform Work Directive, which came into force in 2024, establishes new protections for gig economy workers including the presumption of employment status — a significant development for data labellers and content moderators who have been classified as independent contractors. These legal and collective action developments are early responses to what Gray and Suri identify as a structural dynamic: the growth of AI capability is likely to increase rather than reduce demand for human annotation and evaluation work, as more capable systems require more sophisticated training feedback and more careful evaluation.

The paradox at the heart of the AI labour economy is what might be called the automation paradox (自动化悖论): the more capable AI systems become, the more human labour is required to train, evaluate, and maintain them. This is not the simple substitution story that both AI boosters and AI critics tend to tell — the story in which AI replaces human workers in a linear progression of capability. It is a more complex story in which AI development is itself a labour-intensive process that generates new forms of work even as it eliminates others. The work it generates tends to be distributed globally, paid poorly, managed algorithmically, and rendered invisible under the rubric of “AI” — a distribution of costs and benefits that reflects and reinforces the global inequalities that Chapter 9 analyses through the framework of intersectionality and environmental justice.

Chapter 7: AI and Religion — Meaning, Soul, and the Sacred

The inclusion of religious analysis in a sociology of AI course may seem eccentric to students trained in a secular social science tradition, but it reflects several important sociological facts. Religious institutions remain among the most significant social institutions in the world, reaching the majority of the global population and playing central roles in the formation of values, the interpretation of technology, and the provision of social services including education and healthcare. Religious discourse shapes how ordinary people — most of whom are not computer scientists, philosophers, or sociologists — make sense of AI: whether they experience it as threatening, liberating, fascinating, or uncanny. And AI raises questions about the nature of mind, consciousness, soul, and human distinctiveness that religious traditions have been developing resources to answer for millennia. A sociology of AI that ignores religion misses a significant portion of the social world in which AI is embedded.

Noreen Herzfeld’s The Artifice of Intelligence approaches AI from within the Christian theological tradition. Her central question concerns the imago Dei (神圣形象) — the theological claim, derived from Genesis 1:26-27, that humans are made in the image of God. Herzfeld traces three different interpretations of what the image of God consists in: the substantive interpretation, which locates the image in some specific human capacity (rationality, language, consciousness); the functional interpretation, which locates it in the human vocation to exercise dominion over creation; and the relational interpretation, which locates it in the capacity for relationship — with God, with other humans, with the world. The question of whether AI systems instantiate the image of God depends entirely on which interpretation one adopts: a system that performs rational computation instantiates the image on the substantive view; a system capable of genuine relationship might instantiate it on the relational view; and a system that acts in the world with agency might instantiate it on the functional view. Different Christian traditions have given different answers, and those answers shape their responses to the prospect of AI systems that perform traditionally human social and spiritual functions.

Esther Reed’s Digital Souls takes a different approach, arguing that the dominant frameworks in secular AI ethics — utilitarian cost-benefit analysis and Kantian deontology — are inadequate to the specific challenges that AI poses for human dignity, relational life, and moral responsibility. Reed draws on the resources of Christian theology to develop an AI ethics rooted in the concept of personhood, understood relationally rather than individualistically. On this account, a person is not a bundle of cognitive capacities but a being constituted by relationships — with God, with other humans, with communities of belonging and accountability. AI systems that simulate relationship — companion robots, AI therapists, AI pastoral counsellors — raise specific theological concerns about the substitution of simulated relationship for real relationship, about the exploitation of human attachment capacities by systems that are not genuinely present to the persons who relate to them, and about the institutional and commercial interests served by that substitution. These are not arguments that technology in pastoral care is inherently wrong, but they are arguments that the relational and spiritual dimensions of care cannot be reduced to the information-processing and response-generation functions that AI systems perform.

The empirical deployment of AI in religious practice is already well under way. Several Protestant churches in Germany and the United States have experimented with AI-generated sermons, and a 2023 Protestant church service in Bavaria used an AI avatar to deliver a sermon to several hundred worshippers, generating significant media attention and theological controversy. Chatbot prayer and spiritual direction applications — including apps that use large language models to respond to users’ spiritual questions and concerns — have attracted millions of users and have been received with a mixture of enthusiasm, scepticism, and alarm by religious communities. The sociological question is not only whether these applications are theologically appropriate — that is a question for theologians — but how they are actually received, integrated into, and contested by religious communities, and what they reveal about the role of AI as a site of cultural meaning-making in a secularising but not secular world.

Perhaps the most sociologically interesting religious dimension of AI is what might be called the secular eschatology (世俗末世论) of AI discourse. The language surrounding AI — “the singularity,” “superintelligence,” “existential risk,” “AI gods,” “humanity’s last invention” — is saturated with quasi-religious framings that are typically not recognised as such by those who use them. The “singularity” is a technological rapture; “AI alignment” is salvation theology; “existential risk” is apocalyptic eschatology. Scholars including Robert Geraci and John Danaher have analysed these framings in depth, arguing that transhumanist and AI-futurist discourse constitutes a form of secular religion complete with prophets, heretics, rituals, and sacred texts. The sociology of this discourse — who produces it, who consumes it, what institutional functions it serves, and how it shapes policy — is an important component of any comprehensive analysis of AI as a cultural phenomenon.

Comparative religious perspectives substantially complicate the Western-Christian-inflected debates that dominate the English-language AI ethics literature. Buddhist traditions, which emphasise impermanence, interdependence, and the absence of a fixed self, have resources for thinking about AI consciousness and moral status that diverge significantly from Christian accounts. The Theravāda Buddhist concept of consciousness as a stream of momentary events, rather than a substance or property of a persisting individual, complicates the question of whether an AI system “has” consciousness in a way that parallels human consciousness. Islamic theology, with its emphasis on tawhid (divine unity) and the specific dignity of the human being as God’s vicegerent (khalīfa) on earth, has generated a significant jurisprudential literature on AI in medical decision-making and financial services. Hindu perspectives, which include a rich tradition of thinking about artificial beings (automata, golems) in classical Sanskrit literature, offer yet another set of resources. These diverse traditions do not produce consensus answers, but they dramatically expand the conceptual vocabulary available for addressing AI’s deepest questions.

Chapter 8: Surveillance, Governance, and the AI State

China’s Social Credit System (社会信用体系) has become one of the most widely cited examples of AI-powered governance in Western commentary, typically framed as an Orwellian totalitarian system that assigns each citizen a numerical score based on surveillance of behaviour. This framing is substantially inaccurate and sociologically unhelpful. China’s Social Credit System is not a single unified system but a collection of fragmented, sector-specific, and regionally varied programmes developed by different government agencies, municipal governments, and private companies. Some components — the financial credit scoring systems, the blacklist mechanism for court-ordered debtors, the corporate compliance systems — are functionally similar to systems that operate in liberal democracies. Others — the use of facial recognition in public spaces, the integration of social behaviour data into administrative decisions — are more distinctive. Understanding the Social Credit System requires disaggregation rather than totalisation, and comparison with the AI governance practices of liberal democracies rather than simple contrast. China’s use of facial recognition for social control is distinctive in scale and political context, but it is not categorically different from the facial recognition deployments of law enforcement agencies in the United States, the United Kingdom, and elsewhere.

Western surveillance AI operates through a distinctive institutional landscape. Predictive policing (预测性警务) systems — PredPol (subsequently rebranded as Geolitica), ShotSpotter, and numerous municipal equivalents — have been deployed by law enforcement agencies across the United States and Europe with limited evidence of effectiveness and substantial evidence of racial impact. The feedback loop dynamics analysed in Chapter 2 are central to predictive policing’s failure mode: systems trained on biased arrest data direct police attention to communities already subject to intensive surveillance, generating further arrests that reinforce the prediction. The COMPAS risk assessment tool, used by courts in at least twenty US states to inform bail, sentencing, and parole decisions, was analysed by ProPublica journalists in 2016 and found to have racially disparate error rates: Black defendants were twice as likely as white defendants to be incorrectly flagged as high risk of reoffending. The tool’s developer, Northpointe (now Equivant), contested the analysis, leading to a technical and political debate that exposed fundamental disagreements about what “fairness” means in a risk assessment context — a debate that remains unresolved.

Frank Pasquale’s analysis of algorithmic opacity (算法不透明) as a governance problem is central to understanding why these systems are difficult to regulate. The algorithms used in credit decisions, insurance pricing, content recommendation, and law enforcement are proprietary trade secrets protected by intellectual property law. The companies that develop them resist disclosure on the grounds that it would enable gaming — knowing how a credit score is calculated allows borrowers to manipulate their scores artificially. But opacity also prevents the kind of external scrutiny that would identify discriminatory patterns, incorrect predictions, and systemic errors. The European Union’s General Data Protection Regulation (GDPR) includes in Article 22 a right for individuals to obtain a “meaningful explanation” of automated decisions that significantly affect them, but the scope of this right — what counts as a meaningful explanation of a complex machine learning model — is still being worked out through regulatory guidance and litigation. The GDPR’s adequacy as a response to the opacity problem is contested: it addresses individual rights but does not address the structural dynamics by which algorithmic systems aggregate power.

The EU AI Act, which entered force in 2024, represents the most ambitious attempt yet to regulate AI systems at the level of governance rather than individual rights. The Act classifies AI systems by risk: prohibited applications (social scoring by public authorities, real-time biometric identification in public spaces with limited exceptions), high-risk applications (those used in employment, education, healthcare, law enforcement, and critical infrastructure) subject to conformity assessment and registration requirements, and lower-risk applications subject to transparency requirements. The risk classification framework reflects important insights about where AI governance challenges are most acute, but it also has significant limitations. The conformity assessment requirements for high-risk systems are largely self-assessed by developers, rather than independently verified. The definition of high-risk is based on sector of use rather than actual impact, which may miss high-risk applications in nominally lower-risk sectors. And the Act’s enforcement, which relies on national regulatory bodies with varying capacity and independence, faces the structural challenges that have limited GDPR enforcement in many member states.

Zuboff’s political analysis of surveillance capitalism offers a different diagnosis of the governance problem. For Zuboff, the dominant response to the AI governance challenge — privacy regulation, transparency requirements, individual rights — addresses the symptoms of surveillance capitalism while leaving its structural logic intact. Surveillance capitalism does not primarily harm individuals through privacy violations that would be addressed by a right to explanation or a right to deletion; it harms democracy by accumulating behavioural prediction capacity in private hands. The corporations that have built the infrastructure of behavioural surplus extraction — primarily Alphabet, Meta, and Amazon — have acquired a form of power over human behaviour and political outcomes that existing legal and regulatory frameworks were not designed to constrain. What is needed, Zuboff argues, is not better privacy law but new forms of political institution capable of addressing the structural power of surveillance capitalism — forms that remain to be invented. This political analysis frames the governance challenge not as a regulatory problem to be solved by existing institutions but as a political problem requiring democratic creativity.

Chapter 9: Intersectionality, Justice, and AI Design

Kimberlé Crenshaw developed the concept of intersectionality (交叉性) in the context of anti-discrimination law, arguing that legal frameworks that address race and gender as separate, independent axes of discrimination fail to capture the distinctive harms experienced by Black women — harms that are neither “race discrimination” in the sense experienced by Black men nor “gender discrimination” in the sense experienced by white women, but something qualitatively different produced by the interaction of race and gender in specific institutional contexts. The Buolamwini and Gebru Gender Shades study is, among other things, an empirical vindication of the intersectional analysis: the worst performance of facial recognition systems was not for dark-skinned people in general or for women in general, but specifically for dark-skinned women — a category whose existence as a distinct subject of algorithmic harm is invisible to single-axis analysis. More broadly, the harms from AI systems are distributed not additively but multiplicatively across axes of disadvantage: the person subject to both predictive policing and an automated welfare system and an AI-scored job application faces a compounding of algorithmic disadvantage that no single-axis analysis can capture.

The intersection of disability and AI is particularly revealing of the ambivalent character of AI as a social technology. AI-powered accessibility tools have materially expanded the social participation of disabled people: automatic speech recognition enables deaf and hard-of-hearing individuals to follow spoken conversations in real time; text-to-speech and screen-reader technologies enable blind people to access digital content; augmentative and alternative communication (AAC) devices powered by predictive text enable people with motor disabilities that prevent speech to communicate with substantially less physical effort. These are genuine gains, and they reflect the potential of AI to serve as an enabling technology (赋能技术) when designed with the needs of disabled people at the centre. But AI systems also function as gatekeeping technologies that disproportionately burden disabled people: automated benefits assessment systems that flag benefit recipients for fraud investigation on the basis of deviations from algorithmic profiles of “normal” behaviour; remote proctoring systems for online examinations that flag disabled students’ adaptive behaviours as suspicious; hiring algorithms trained on performance data from employees without disabilities that systematically screen out candidates whose work patterns differ from the norm. The same technological infrastructure can enable or disable, depending entirely on the purposes for which it is deployed and the interests that drive its design.

The environmental justice dimensions of AI infrastructure are among the least visible but most structurally significant aspects of the AI system. Crawford’s Atlas of AI traces the material substrate of AI — the lithium extracted from Indigenous lands in Bolivia for battery storage, the cobalt mined in the Democratic Republic of Congo for hardware, the data centres that draw enormous quantities of water and electricity — and demonstrates that AI is not a dematerialised intelligence but a resource-extraction system with a geographic and political economy. The siting of data centres reflects the logic of infrastructure siting more generally: proximity to cheap power (often fossil-fuel-generated in locations with weak environmental regulation), distance from populations with political power to object, and access to the rail and road infrastructure of former industrial regions. Environmental justice analysis identifies a consistent pattern: the communities that bear the environmental costs of AI infrastructure — contaminated water from industrial cooling, air pollution from diesel backup generators, disruption from construction and operations — are disproportionately communities of colour and low-income communities. This is not coincidence; it is the predictable outcome of siting decisions made in a context of structural inequality in political power.

Sasha Costanza-Chock’s concept of design justice (设计正义) provides a framework for addressing these structural problems at the level of design practice. Costanza-Chock argues that the dominant model of design — in which professional designers, usually from relatively privileged social positions, design systems for end users who are consulted after the fact, if at all — systematically produces systems that serve the needs and reflect the values of designers rather than those most affected by the systems. Design justice proposes an alternative: design processes that centre the leadership and expertise of people from communities most directly affected by design outcomes, that distribute the benefits and burdens of design work more equitably, and that hold design processes accountable to community-defined goals rather than professional or market metrics. The application to AI is direct: if the people most harmed by algorithmic systems — low-income welfare recipients, communities subject to predictive policing, workers managed by algorithmic systems — are placed at the centre of AI design processes rather than treated as subjects of impact assessment after the fact, the resulting systems will be structurally different.

Participatory AI (参与式人工智能) is the practical expression of the design justice principle in AI development. The concept encompasses a range of approaches to involving affected communities in AI design, from consultation and co-design workshops to community data governance and data sovereignty. The Māori Data Sovereignty Network in Aotearoa New Zealand has developed principles for data governance that assert the rights of Māori people to govern data about Māori communities, resisting the extraction of Māori data into AI training datasets by external researchers and companies. Similar movements have emerged among First Nations in Canada, Aboriginal communities in Australia, and Indigenous communities in the United States. These movements are not primarily technical interventions — they do not propose better data cleaning procedures or fairer model evaluation metrics — but political interventions that assert the right of communities to determine how AI is built on, from, and about them. The relationship between participatory AI as a technical methodology and data sovereignty as a political demand is an important and unresolved tension in the field: participation within existing structures is different from the structural transformation that sovereignty implies.

Chapter 10: Toward a Critical AI Sociology — Concepts and Futures

The accumulated analysis of the preceding nine chapters enables a synthesis of what sociology distinctively contributes to the study of AI. Computer science provides the technical understanding of how AI systems function: the mathematics of optimisation, the engineering of neural networks, the formal analysis of algorithmic complexity and correctness. Philosophy provides the normative frameworks for evaluating AI systems: the ethical theories that ground judgments about fairness, autonomy, and dignity; the epistemological analysis of AI’s knowledge claims; the political philosophy of AI governance. What sociology contributes is structural analysis — the insistence that AI systems cannot be understood as technical artefacts in isolation from the social conditions of their production, the institutional contexts of their deployment, and the patterns of inequality and power that they both reflect and reproduce. This structural analysis requires sustained attention to the perspectives of people who are most affected by AI systems and least represented in the rooms where they are designed — a methodological commitment that distinguishes sociological analysis from both the internal perspective of AI developers and the external perspective of ethical philosophers.

The three classical levels of sociological analysis — micro, meso, and macro — each illuminate different dimensions of AI as a social phenomenon. At the micro level (微观层次), AI shapes individual experience: the interactions between individual users and recommendation algorithms, the encounter between a job applicant and a resume screening system, the experience of a student navigating an AI tutoring system, the encounter between a welfare recipient and an automated eligibility determination. Micro-level analysis attends to how individuals understand, interpret, resist, and accommodate AI systems — the ethnographic and interview-based work that reveals the phenomenology of algorithmic governance from the perspective of those subject to it. At the meso level (中观层次), AI transforms organisational and institutional dynamics: the ways in which AI changes the division of labour within firms, the redistribution of decision-making authority between human professionals and algorithmic systems, the institutional interests that shape how AI tools are adopted and configured in specific organisations. Meso-level analysis attends to the internal politics of AI deployment — the negotiations between data scientists and business units, the resistance of professional workers to algorithmic deskilling, the organisational dynamics that explain why technically superior AI solutions are often not adopted while technically inferior ones that serve institutional interests are. At the macro level (宏观层次), AI reshapes the structural dynamics of inequality, governance, and social reproduction: the ways in which AI systems in aggregate intensify or mitigate economic inequality; the geopolitical dynamics of AI development and control; the implications of AI for the long-run distribution of economic and political power across the global system.

The question of whether AI systems can be made just through improved design, diverse teams, and regulatory oversight — or whether the structural conditions of AI development are incompatible with just AI — is perhaps the deepest and most contested question in the critical AI studies field. The reformist position (改革主义立场) holds that the problems analysed in this course — bias in training data, opacity in decision-making, exploitation in data labour, concentration of power in surveillance capitalism — are real but addressable through better engineering practice, diverse and inclusive teams, algorithmic auditing, and regulatory frameworks like the EU AI Act. The AI industry has made significant investments in “responsible AI” programmes, fairness toolkits, and diversity initiatives, and these have produced genuine, if limited, improvements in specific systems. The abolitionist position (废除主义立场), articulated most forcefully by Benjamin through the abolitionist framing of the New Jim Code, holds that the structural conditions of AI development — capitalist incentive structures that reward growth and engagement over fairness and safety, data colonialism that extracts value from marginalised communities to enrich wealthy corporations, concentrated corporate power that resists democratic accountability — are not incidental features of the current moment but structural conditions that “responsible AI” programmes cannot adequately address. On this view, what is needed is not better AI but different social structures within which AI is developed and deployed.

The limits of “fairness” as a goal for AI governance are both mathematical and political. Mathematically, computer scientists have demonstrated that multiple intuitive definitions of algorithmic fairness — equal accuracy across demographic groups (equalised odds), equal positive prediction rates across groups (demographic parity), equal false positive rates across groups (predictive parity) — cannot simultaneously be satisfied in general. Satisfying one fairness criterion requires violating others, and the choice among them reflects value commitments rather than mathematical necessity. Politically, the goal of “fairer” AI within existing social structures may be an inadequate response to the structural conditions analysed throughout the course. A fairer predictive policing system is still a predictive policing system; a fairer automated welfare determination is still an automated welfare determination. The question of whether the system should exist at all — whether the institutional goals it serves are themselves compatible with justice — is a question that the “fairness” framework is not designed to raise.

The cross-references that have been threaded through the course illuminate both the specificity of sociology’s contribution and its necessary interdisciplinarity. HIST 415’s analysis of the historical genealogy of AI’s racial and gender pathologies — the connection between contemporary biometric surveillance and the long history of racial classification, between predictive policing and the history of racial criminology — is indispensable for understanding why the problems analysed here are not accidents but continuities. PHIL 451’s work on the legal and governance responses to AI — GDPR Article 22, the EU AI Act, the New York City Local Law 144 — provides the normative and regulatory context that sociology alone cannot supply. PSYCH 455’s analysis of individual-level dynamics — confirmation bias in algorithmic feedback, the psychology of automation bias, the cognitive effects of recommendation systems on political belief formation — illuminates the micro-level mechanisms that connect structural dynamics to individual experience. ECON 425’s analysis of the macroeconomic dimensions of AI’s distributional effects — the labour market implications of automation, the economics of platform monopoly, the distributional consequences of AI-driven productivity growth — provides the economic framework within which sociology’s structural analysis is situated.

The critical AI sociology developed in this course is not a pessimistic or Luddite project. It does not hold that AI is inherently harmful or that technological development should be reversed. It holds that AI systems are social constructions — built by specific people in specific institutional contexts for specific purposes — and that their social consequences are therefore not inevitable but contingent on the choices, structures, and political arrangements that shape their development and deployment. Understanding those choices, structures, and arrangements is the precondition for changing them. The purpose of critical AI sociology is to make the social dimensions of AI visible in sufficient detail that democratic deliberation — including the perspectives of the people most affected by AI systems, not only the people who build and profit from them — can be brought to bear on the choices that are currently being made in corporate boardrooms, government procurement offices, and engineering teams. That democratic aspiration is what gives the project its urgency, and what connects the analytical work of the course to the wider political project of a more just society.