Henry Farrell, Alison Gopnik, Cosma Shalizi and James Evans (2025), Science 387(6739):1153–1156, Policy Forum — DOI-10.1126/science.adt9819. Read from the authors’ accepted version.
Key takeaways
- The thesis is a classification, not a prediction. 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.” The question of their effects is then asked separately, and the paper insists that calling them cultural technologies raises, not lowers, their likely impact.
- Two lineages combined. As cultural technologies (pictures, writing, print, video, search) they give access to information others created. As social technologies, in Simon’s sense of “artificial systems of human society” (markets, bureaucracies, democracies), they also reorganize it. Prices, election results, GDP and bureaucratic categories are “lossy (i.e., incomplete, selective, and uninvertible) but useful representations”; a large model is a lossy JPEG of its corpus in the same sense (Chiang’s phrase).
- The case against agency is short. Agency here means perceiving and acting on a changing world, building and revising models of it and setting novel goals. Large models do not do this: hallucination is endemic because they “have no conception of truth and falsity”, and scaling “won’t change this”. This is asserted, not argued.
- Why we see agents. Stories pass on information through fictional agents that listeners know are not real; “Chatbots are the successor to Hercules, Anansi, and Peter Rabbit”, and markets and states are likewise treated as agents. Behind the interface the human contribution is everywhere: corpora, RLHF raters, prompt engineers, users. A cover-letter request is “a technically mediated relationship with thousands of earlier job applicants”.
- One stated mechanism. Models reproduce sequence probabilities well on average, so they are most accurate where data are common and least where rare or novel, which might worsen Weberian homogenization. Against this: deliberately diversified “society-like” ecologies of models might keep or recombine perspectives. Neither is tested.
- Political economy. The old tension between producers and distributors of information is sharpened by centralized ownership; countervailing institutions (editors, libel law, deposit insurance, the SEC) did not arise on their own but from sustained effort.
Standing in the record
Filed on 2026-10-09 at the owner’s direct request. It is not one of the
works from the bibliography of the manuscript what-survives-translation;
the owner asked for it by itself.
The record’s cluster on LLM selfhood reads large models through the
simulator/simulacra account (LIT-460) and Shanahan’s later papers (LIT-465,
LIT-466). This paper declines that question and sets the models beside
markets and bureaucracies, which is why it is held here with
social-science first. Its explanation of why chatbots look like agents
(fiction’s illustrative agents) is close to LIT-460’s role play but is
reached from the other side: LIT-460 starts from the model and finds a
character; this paper starts from the culture and finds a storytelling
device.
An anthology topic, deployment-and-society, could hold it, and it has
programmatic remarks for AI practice (“thinking in this way might reshape
AI practice”), so it carries anthology-candidate. It is kept here
because its argument is social theory.
Read on 2026-10-09 (NOTE-649). The reading is the source of THEORY-173, Proposed: the paper offers a classification and no test of it.