"AI as Social Technology" is an academic essay by Henry Farrell (Johns Hopkins SAIS) and Cosma Rohilla Shalizi (Carnegie Mellon), published by the Knight First Amendment Institute (Knight Columbia) / Knight Columbia on May 11, 2026. It argues that large language models are best understood as social technologies — systematic means of reorganizing social relationships among human beings — rather than as nascent autonomous agents, and it introduces "social technology" as a named analytical category for LLMs. The essay anchors the Knight Columbia 2026 AI-in-democratic-society symposium.
Central thesis
Farrell and Shalizi argue that LLMs are best understood not as nascent autonomous agents (the Singularity / AGI framing) but as social technologies, continuous with bureaucracies, markets, and democracy (Herbert Simon's "sciences of the artificial") rather than with science-fiction visions of paperclip maximizers or omniscient bureaucracies of terror.
The essay builds on collaborative work with Alison Gopnik and James Evans (Farrell et al., 2025) — the "AI as Cultural Technology" thesis — but emphasizes the social rather than the cultural aspects of LLMs.
Coarse-grainings
The central analytical move is the concept of a coarse-graining: a stripped-down representation of a complex phenomenon that captures key aspects and dynamics by discarding most details. The authors hold that coarse-grainings are ubiquitous, since no scientific model, organism, or artifact grasps the full detail of its environment. Their examples include:
- Macaque subordination signals tracking power relations
- Bureaucratic standards, statistics, and census categories
- The price mechanism (Hayek's "computational limits of human beings")
- Opinion surveys and censuses
- Platform-company embeddings for content recommendation
- LLMs (as coarse-grainings of vast textual corpora)
On this account, all large-scale social institutions process information by reducing complex realities into more tractable abstractions, and economic, administrative, and political coordination at scale is held to be impossible without compressing complex social relationships into visible, tractable representations.
The Long Industrial Revolution framing
The essay dates the "Singularity" not to 1993 (Vinge) but to the Industrial Revolution two centuries ago. The modern social sciences emerged from those shocks, and Farrell and Shalizi argue they now need to work with computer science to map what they describe as another stage in the Long Industrial Revolution. AI may turn out to be very important, they write, but in quite different ways than inherited myths suggest.
This framing engages with AI as Normal Technology (Narayanan-Kapoor) while extending it. Where Narayanan and Kapoor emphasize the diffusion and applications lag that makes AI continuous with prior general-purpose technologies, Farrell and Shalizi emphasize the institutional-organizational dimension — the way Industrial Revolution social technologies (bureaucracies, markets, the price mechanism, print capitalism, opinion surveys) reorganized social relationships at large scale via coarse-grainings.
Lossiness and power
The essay develops two dimensions along which coarse-grainings are compared: lossiness and power.
On lossiness, coarse-grainings necessarily discard information. The authors cite Maxim Raginsky's framing that "abstraction hides a great deal of complexity from view, and this is both its main virtue and its primary peril." They note that mathematical studies of coarse-graining (Chorin et al. 2000; Crutchfield-Feldman 2003) show repressed aspects at best return as statistical noise and at worst as systematic errors. Statistical models trained for good on-average performance trade worse performance in rare situations for better performance in common ones, with implications for small groups, under-represented populations, and novel situations.
On power, coarse-grainings get embroiled in power relations because abstractions create winners and losers, and different simplifications advantage different groups. The authors apply James Scott's Seeing Like a State logic — that bureaucratic simplifications may reshape social organization in their image or create pushback — to the simplified representations of "high tech modernism" (Farrell-Fourcade 2023). They note that platform-company embeddings reshape user self-understandings (Fourcade-Healy 2025).
LLMs create social relations
The essay argues that LLMs create social relations between users and the authors of training-corpus text:
"LLMs create social relations between their users and the authors of the text in their training corpora. With the right access to the model and the corpus, one can trace the connections from system output back to individual source texts and their authors (Grosse et al., 2023). These social relations are mechanically mediated, giving users the illusion that they are interacting with just the machine and not an assemblage of people."
This positions LLMs alongside libraries and languages (cultural transmission) and markets and bureaucracies (institutional coordination). The authors describe the illusion of unmediated machine interaction as a common fact of modern life, with the novel element being the mechanical mediation of relations between a user and an entire body of human culture via a generative coarse-graining.
Engagement with AGI, Singularity, and AI x-risk discourse
The essay engages directly with AGI, Singularity, and AI existential-risk arguments:
- On paperclip maximizers, omniscient bureaucracies of terror, and self-aware markets, the authors argue these visions are not too weird but not nearly weird enough: the possible futures, they write, are much messier and more varied than stark AGI visions, and will be shaped by the collision between imperfect and highly complex technologies and imperfect and highly complex human social systems.
- On Marc Andreessen's "alchemy" / "Philosopher's Stone" framing, the essay treats it as paradigmatic of speculative non-fiction that elides institutional complexity, citing Andreessen (2023) explicitly.
- On the Singularity-to-AGI lineage, the essay traces it through Vinge (1993) to modern progenitor and funder embrace (Becker 2025; Hao 2025), writing that "it is a scandal that this dream of the '90s is still alive and shaping debate."
- On the institution-as-intelligent-agent metaphor, the essay holds that bureaucracies, nations, and markets have historically been anthropomorphized (Lovecraft's racist-conservative framing; Jarrell's humanist critique; Scott's anarchist eulogies), and that current AI metaphors borrow from these prior debates.
The bureaucracy-AI collision
The essay's most policy-relevant claim is that arguments about AGI were "seemingly one influence on the Trump administration's sweeping cutbacks to the administrative state." Farrell and Shalizi argue this is a category error: AI is a new social technology that will interact with bureaucracy and other older social technologies, not a replacement that obsoletes the need for institutional coordination. On their framing, AI does not eliminate the need for bureaucratic coordination but changes the coarse-grainings, with their own (different) attentional trade-offs.
This argument connects to Intelligence Replaces Hierarchy (the Guardian May 15 piece on manager purges; the Coinbase/Block 1:175 spans; Meta async management), AI Labor Disruption (the Economist jobs-apocalypse cluster, May 14), and the DOGE — Department of Government Efficiency (AI Deployer) administrative-state cuts.
The research-agenda call
The essay closes with a programmatic call for collaboration between social scientists and computer scientists:
"We urgently need to discover how the new coarse-grainings of AI interact with the existing abstractions through which humans simplify an inherently complex world to make it tractable. Both AI and older social technologies are, among other things, forms of information processing. We should investigate how the former are variously reinforcing, reshaping, or replacing the latter."
It cites DeDeo (2017): "Once we realize that the machine-aided predictors of a system are also participants, it is natural to ask how their use of that knowledge, accurate or not, back-reacts on the society itself."
Relation to adjacent frameworks
The essay names "social technology" as a positive analytical category for LLMs, distinct from but complementary to two adjacent frameworks:
- Cultural technology (Yiu-Kosoy-Gopnik 2024; Farrell-Gopnik-Evans-Shalizi 2025) — emphasizes cultural transmission.
- Normal technology (Narayanan-Kapoor 2025) — emphasizes continuity with prior general-purpose technologies.
- Social technology (Farrell-Shalizi 2026) — emphasizes the coarse-graining and institutional-coordination dimension.
The authors present all three as complementary, each adding explanatory power the others lack. Because the essay introduces a named framework that can be cited by author, it is classified as a foundational source.
Relationships
- supports: AI as Social Technology, AI as Social Technology, The Long Industrial Revolution
- engages-with: AI as Normal Technology (Narayanan-Kapoor 2025) — complementary; emphasizes institutional dimension Narayanan-Kapoor under-emphasize
- engages-with: AI as Cultural Technology (Farrell-Gopnik-Evans-Shalizi 2025; Yiu-Kosoy-Gopnik 2024)
- contradicts: Artificial General Intelligence (AGI), Superintelligence, Technological Singularity framings of LLMs as nascent autonomous agents (treats the framing as a category error, not as false — the disagreement is structural, not factual)
- related: Intelligence Replaces Hierarchy (the bureaucracy-AI collision angle), AI Political Economy, AI and Democracy, Seeing Like a State (applied to AI) (James Scott), Henry Farrell, Cosma Rohilla Shalizi, Alison Gopnik, Herbert A. Simon
- part-of cluster: Knight Columbia 2026 AI-in-democratic-society symposium
Sources
Raw Sources/AI as Social Technology.md- Published at Knight Columbia: knightcolumbia.org