Are large AI models actually intelligent agents or social technologies?

Farrell, Gopnik, Shalizi, and Evans argue that "the discourse about large models as intelligent agents is fundamentally misconceived." Their claim is that Large Models "should not be viewed primarily as intelligent agents, but as a new kind of cultural and social technology, allowing humans to take advantage of information other humans have accumulated." They place Large Models in a lineage running through "writing, print, markets, bureaucracies, and representative democracies" — technologies that let one person benefit from information accumulated by many others, at scale.

The mechanism they give is the "lossy representation." Market prices, bureaucratic categories, and census statistics each compress enormous amounts of distributed human knowledge, value, and action into a usable but incomplete summary — "a price... compresses vast swathes of detail into a simplified but usable representation." Large Models do the same to cultural material: they are "lossy JPEGs" of their training corpora, generating "summaries of unmanageably large and complex bodies of human-generated information" and then letting that compressed representation be "reorganized and reconstructed" in new ways — condensing prose, inflating it into metaphor, generating images or video from a prompt. Crucially, the authors note the systems still depend on continuous human input to function: RLHF raters, prompt engineers, and users who begin "even the newest 'chain of thought' models" with dialogue. On their account, "hallucination" is not a bug to be scaled away but a structural consequence of having "no conception of truth and falsity" — the system samples and generates, it does not know.

This sits close to Should restricting AI access create new kinds of inequality?, which also treats Large Models as processors of aggregated human output rather than independent minds. Farrell et al. go further than that framing: rather than asking who owns the aggregate (the copyright question), they ask what kind of thing an aggregator-that-reorganizes is, and place it next to markets and bureaucracies specifically because those institutions also generate lossy summaries that then feed back and reshape the individual knowledge and decisions that produced them. That feedback loop — social technologies shaping the individuals who feed them — gives Does incremental AI replacement erode human influence over society? a historical precedent: central banks' narrow financial variables "provided apparent financial stability" while letting housing-market instability build up unseen, which the authors offer as a caution about what any lossy institutional summary can hide. It also anticipates What makes accountable judgment scarce when AI cognition is cheap?: Farrell et al.'s own examples of print, markets, and bureaucracy argue that it was institutions, not the raw technology, that determined who gained and who lost from earlier general-purpose information technologies.

The paper is an argued reframing, not a measurement — it offers no data on whether current AI deployment is in fact producing the homogenization, fragmentation, or labor displacement it flags as open questions ("Will Large Models homogenize or fragment culture and society?" is explicitly left unresolved, with Weber's bureaucratic-rationalization worry and Mill's "doux commerce" optimism offered as the two historical poles without a verdict between them). What follows at the strength the excerpt actually supports is narrower than "AI will reshape society like print did": it is that the right unit of analysis for large models is institutional and historical rather than agentive, and that the specific social-science toolkit built to study markets, bureaucracies, and media is the toolkit this debate is missing.