AI political economy is an analytic lens that treats AI, and large language models in particular, as a force that reorganizes social relationships and redistributes economic, informational, and political power, rather than as a nascent autonomous intelligence. It frames AI deployment as a question of who gains and who loses, and is anchored most directly on the Farrell–Shalizi "AI as social technology" framework (AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026)).
Core claim: power flows through coarse-grainings
The framework's central claim is that all large-scale social institutions coordinate by compressing complex reality into tractable abstractions, which Farrell and Shalizi call coarse-grainings. Markets compress dispersed knowledge into prices; bureaucracies into standards, statistics, and census categories; democracies into votes and opinion surveys; platforms into embeddings. LLMs are the newest such coarse-graining, a lossy compression of a vast textual corpus (AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026)).
Two properties make coarse-grainings inherently political. The first is lossiness: every abstraction discards information, and repressed detail returns "at best as statistical noise; at worst as systematic errors," falling hardest on small groups, under-represented populations, and novel situations (see AI Bias and Discrimination, Mathematical Impossibility of Perfect Fairness). The second is power: because different simplifications advantage different groups, abstractions create winners and losers. Whoever controls the coarse-graining, including which census categories exist, which content a recommender surfaces, and which corpus a model is trained on, shapes the distribution of advantage. Farrell and Shalizi present this as the Scott "seeing like a state" logic applied to high-modernist AI (Farrell–Fourcade 2023; Fourcade–Healy 2025).
This reframes AI policy questions as distributional and institutional rather than purely technical or existential: the question is less "will the model escape?" and more "who governs the abstraction, and on whose behalf?" The framing connects to AI Power Concentration (the who-governs risk) and is distinguished from the loss-of-control framing of AI Existential Risk.
Relationship to adjacent frameworks
AI political economy sits within a family of "demystifying" frameworks:
- AI as Social Technology (Farrell–Shalizi) — the parent framework, emphasizing the institutional-organizational dimension (how AI reorganizes relationships at scale).
- AI as Cultural Technology (Farrell–Gopnik–Evans–Shalizi) — emphasizing cultural transmission (AI as akin to libraries and language).
- AI as Normal Technology (Narayanan–Kapoor) — emphasizing diffusion and applications lag.
- The Long Industrial Revolution — the historical framing, treating AI as another stage in a two-century reorganization of social technologies (bureaucracies, markets, print capitalism) rather than a 1993-style "Singularity."
Where the "normal technology" view stresses continuity through slow diffusion, the political-economy view stresses continuity through the politics of abstraction: on this account AI may matter greatly, but in the way that the price mechanism or the census mattered, by changing who sees, who decides, and who is made legible to whom.
The framing also bridges to the labor and concentration threads documented in AI Labor Disruption, AI Power Concentration, and AI Divides (Literacy / Occupational / Ethico-Philosophical), which are political-economy questions in everything but name.
Contested points
- How novel is it? Critics of the social-technology framing argue it under-weights genuinely new affordances, such as open-ended generation and agentic action, by assimilating LLMs too neatly to markets and bureaucracies. Defenders reply that the burden of proof runs the other way, on the grounds that "this time is different" claims about technology usually recapitulate earlier ones.
- Determinism vs. contingency. The framework holds that outcomes are contested rather than technologically determined: the same coarse-graining can entrench or redistribute power depending on who controls it, which is why Farrell and Shalizi cast it as a political economy rather than a technology forecast.
Relationships
- depends-on: AI as Social Technology, Seeing Like a State (applied to AI)
- related: AI as Cultural Technology, AI as Normal Technology, The Long Industrial Revolution, AI Power Concentration, AI Labor Disruption, AI Bias and Discrimination
- instance-of: AI as Social Technology
Provenance note: Built primarily from the ingested AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026) essay (Farrell + Shalizi, Knight Columbia, May 11 2026); single load-bearing source, hence
confidence: medium.