Source
Farrell, Gopnik, Shalizi and Evans (2025), LIT-848, read in NOTE-649 from the authors’ accepted version.
What was actually shown
Nothing was tested. The paper argues by classification and analogy.
- The classification. Like writing, print and search, large models give access to what other people produced. Like markets, bureaucracies and elections (Simon’s “artificial systems of human society”), they make “lossy (i.e., incomplete, selective, and uninvertible) but useful” representations of it that support new operations at scale.
- The denial of agency. Agents perceive and act on a changing world, revise models of it from evidence and set novel goals. Large models sample from an estimated distribution over sequences and have “no conception of truth and falsity”; scaling “won’t change this”. The agent-like interface continues the storyteller’s use of fictional agents.
- The one mechanism. A model fitted to reproduce sequence probabilities on average is most accurate on what is common in its data and least on what is rare or novel. The authors say this “might” worsen cultural homogenization. That the fit is worst on rare cases follows from the training objective; the cultural consequence is not shown.
What could have come out otherwise: nothing in the paper. A demonstrated capacity for evidence-driven revision of a world model in a large model would have counted against the denial of agency, but the paper does not look for one.
What this does not say
- Not that large models are unimportant or harmless. The authors say the classification raises their likely impact, as print’s and markets’ was large.
- Not that no artificial system could be an agent. Robotics and possible future systems are set aside; the claim is about large models as now built and deployed.
- Not that the agent description is useless. The paper says we “tend” to think of large-scale technologies as agents and that this distracts, but markets and firms are productively modelled as agents in economics; the paper does not show that the description misleads more for models than for markets.
- Not a claim about consciousness or selfhood. It is silent on them.
Connections
- The simulator/simulacra account (LIT-460). Compatible, from the other side: there the model role-plays characters with “no-one at home”; here the characters continue fiction’s illustrative agents. Neither attributes goals to the model itself.
- THEORY-023. That account treats the human origin of a model’s output as what defeats behavioural evidence of consciousness. This one treats the same fact as what the model is.
- Repeated reconstruction. The homogenization mechanism is a claim about contraction toward common material when transmission passes through a model fitted to the average. It is stated, not modelled.
Note of 2026-10-29: what holds it back
No criterion that would make a large model an agent rather than a technology has been stated and applied. The account’s distinctive prediction, that routing cultural production through such models narrows what is produced, is untested: THEORY-226 measures a shared direction of LLM revisions, not their dispersion against human revisions. The paper offers no evidence, so the reading holds only as a classification argued by analogy.