AI as social technology is a framework proposed by Henry Farrell and Cosma Rohilla Shalizi in AI as Social Technology (AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026), Knight Columbia, May 11 2026). It argues that large language models and related AI systems are best understood as social technologies — systematic means of reorganizing social relationships among human beings — continuous with markets, bureaucracies, democracy, and the price mechanism, rather than as nascent autonomous agents in the AGI or Singularity tradition.
Core claims
Farrell and Shalizi advance three claims at once. The descriptive claim is that LLMs work as statistical coarse-grainings of vast textual corpora, creating social relations between users and the authors of training-corpus text, mechanically mediated in ways that resemble and interact with existing social technologies. The comparative claim is that AI is the latest stage in the Long Industrial Revolution — the centuries-long transformation of human capacities to organize economic, social, and political life — whose earlier social technologies include bureaucracies, markets, print capitalism, opinion surveys, and democracy itself. The programmatic claim is that social scientists and computer scientists need to collaborate to map how AI coarse-grainings interact with existing abstractions, whether by reinforcing, reshaping, or replacing them.
Coarse-grainings
A coarse-graining is a stripped-down representation of a complex phenomenon that captures key aspects while discarding most details. Farrell and Shalizi treat coarse-grainings as ubiquitous across social institutions: all large-scale social institutions process information by reducing complex realities into more tractable abstractions, a reduction they describe as necessary because no scientific model, organism, or artifact can grasp the full detail of its environment.
| Coarse-graining | Coarse-grained reality |
|---|---|
| Macaque subordination signals | Power relations within the troop |
| Bureaucratic standards and census categories | Heterogeneous populations and social structures |
| The price mechanism | Vast economic systems |
| Opinion surveys | Public opinion |
| Platform-company embeddings | Particular users and the universe of content |
| LLMs | Vast corpora of human textual culture |
Lossiness and power
Farrell and Shalizi compare social technologies along two dimensions: lossiness and power.
On lossiness, coarse-grainings necessarily discard information. They cite mathematical studies of coarse-graining (Chorin et al. 2000; Crutchfield & Feldman 2003) showing that repressed aspects of the process 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, with implications for small groups, under-represented populations, and novel situations. On this account, AI coarse-grainings will have different attentional trade-offs than previous social technologies, but they will have trade-offs.
On power, coarse-grainings become embroiled in power relations because abstractions create winners and losers. The authors invoke James Scott's Seeing Like a State logic — bureaucratic simplifications either reshape social organization in their image or generate pushback — which Farrell and Fourcade (2023) argue broadly characterizes "high tech modernism" as well. Platform-company embeddings, they note, reshape user self-understandings (Fourcade & Healy 2025).
The Simon connection
The framework grounds itself in Herbert Simon's "sciences of the artificial" — AI, political science, administration, economics, computer science, and cognitive psychology as branches of the study of how human beings create "artifacts" that model and act on their environment. From this perspective AI models are another means of complex information processing, encompassing both information technologies built by engineers and social information systems such as markets, bureaucracy, and democracy. Simon's bounded-rationality framework (1957) treats large-scale social technologies as emerging from the need of limited humans to build collective arrangements that allow them to map and manage a complex world; Farrell and Shalizi treat LLMs as the latest entry in this lineage.
Feedback relations
LLMs, social-media recommendation models, and other AI coarse-grainings differ from prior social technologies in their feedback relations with the systems they model. Frontier LLMs are expensive to train and updated only at lengthy intervals, whereas models used for advertising and social-media matching are much cheaper to estimate and change much more quickly. These differences affect the time scales and feedback relationships between coarse-grainings and the social systems they seek to represent (Flack 2017). Hidden prompts allow rapid updates while weight updates remain slow. Farrell and Shalizi argue that a near-future of cheap small models would encourage a less centralized political economy than one in which large models retain a decisive edge, which they frame as a policy-relevant claim.
Engagement with the AI debate
Farrell and Shalizi position their framework against several existing accounts of AI.
| Position | Farrell-Shalizi response |
|---|---|
| AGI / Singularity / superintelligence | A category error rooted in 1990s science fiction; LLMs are not nascent autonomous agents but coarse-grainings of human cultural artifacts. The framings, they argue, are "not weird enough" rather than too weird. |
| AI as Normal Technology (Narayanan-Kapoor 2025) | Complementary, but emphasizes diffusion lag where Farrell-Shalizi emphasize the institutional-organizational dimension. |
| AI as Cultural Technology (Farrell-Gopnik-Evans-Shalizi 2025) | Extends the cultural-technology framing by emphasizing the social (relationship-reorganizing) over the cultural (transmission) aspects. |
| AGI replaces bureaucracy (DOGE administrative-state cuts) | A category error; AI changes the coarse-grainings but does not eliminate the need for institutional coordination. |
| STS critical theory (power relations all the way down) | Insufficient microfoundations. |
| Rationalist / Bayesian-agent strategic competition (the "rationalist" AI safety paradigm) | Microfoundations all the way up; misses how collective phenomena emerge. |
The framework relates to AI as Normal Technology (Narayanan-Kapoor 2025) and AI as Cultural Technology (Farrell-Gopnik-Evans-Shalizi 2025), both of which it complements, and stands against the Artificial General Intelligence (AGI), Superintelligence, and Technological Singularity framings of LLMs as nascent autonomous agents.
Relation to policy
The bureaucracy-AI collision angle connects the framework to policy debates over DOGE — Department of Government Efficiency (AI Deployer) administrative-state cuts and to Intelligence Replaces Hierarchy: Farrell and Shalizi argue that AI does not eliminate the need for institutional coordination but instead changes the coarse-grainings through which institutions operate. The lossiness dimension bears on civil-rights enforcement, since statistical models trade rare-situation performance for common-situation performance, a trade-off relevant to the sociotechnical disclosure regime under Colorado SB 26-189 (2026 — replaces 2024 Colorado AI Act).
Relationships
- supports: AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026) (canonical source)
- complements: AI as Normal Technology (Narayanan-Kapoor 2025) — both reject AGI/Singularity framing; Farrell-Shalizi add the institutional-coordination dimension
- complements: AI as Cultural Technology (Farrell-Gopnik-Evans-Shalizi 2025) — emphasizes the social over the cultural
- contradicts: Artificial General Intelligence (AGI), Superintelligence, Technological Singularity framings of LLMs as nascent autonomous agents
- depends-on: Herbert A. Simon (the "sciences of the artificial" framing), Bounded Rationality (Simon 1957), Seeing Like a State (applied to AI) (Scott 1998)
- related: Intelligence Replaces Hierarchy (the bureaucracy-AI collision angle), AI Political Economy, AI and Democracy, The Long Industrial Revolution, Mechanistic Interpretability
- deploys-in: Policy debates over DOGE — Department of Government Efficiency (AI Deployer) administrative-state cuts (the AGI-replaces-bureaucracy frame)
Sources
- AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026) — Farrell + Shalizi, AI as Social Technology, Knight Columbia, May 11 2026 (canonical anchor)
- Farrell, Gopnik, Evans, Shalizi (2025) — AI as Cultural Technology (predecessor)
- Yiu, Kosoy, Gopnik (2024) — Imitation versus Innovation: What Children Can Do That Large Language and Language-and-Vision Models Cannot (Yet) (foundation)