Simon Goldstein, Harvey Lederman (2025), Inquiry (published online 2026-09-04, doi 10.1080/0020174X.2026.2727584); first circulated as a PhilArchive preprint (GOLWDC-2) — DOI-10.1080/0020174X.2026.2727584
Key takeaways
- On an interpretationist theory of belief and desire, the right subject of LLM psychology is the per-context “instance agent”, not the model. Such instances plausibly have beliefs and desires, and their desires are best described by the “HHH+0” framework: helpful, honest and harmless, plus “zero-shot” intrinsic desires acquired in context.
Seeded from the abstract and a skim, before the reading in NOTE-138. What follows is what the work says about itself.
The authors examine LLMs through interpretationism, the view that an agent has whatever beliefs and desires best predict and explain its actual and counterfactual behaviour. They argue for three claims. The proper object of LLM psychology is the instance agent created at the start of each context, not the underlying model. Given interpretationism, there is a strong case that instance agents have beliefs and desires. Their desires are best captured by the HHH+0 framework: helpfulness, honesty and harmlessness, plus further intrinsic “zero-shot” desires acquired in context. They weigh the main rivals, that LLMs “just” predict the next word and that they merely role-play having attitudes, and discuss the implications for copyright law, AI safety and possible future AI moral status.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 0 seconds of active reading over 0 sessions. It was filed Deferred, unread. NOTE-138 is the close reading of 2026-09-26, and it placed the work: Active — Worth reading as the clearest statement of interpretationist realism about LLM instance agents. It turns “does it want?” into comparisons of rival desire hypotheses against named behavioural results, and it is candid that its central empirical discriminations are untested.
When seeded, before the reading in NOTE-138, the entry gave a priority for a deeper reading: high — t=0, but it is directly load-bearing for LLM agency and welfare notes. The skim captures the architecture but not the §5 evidence or the §7 practical arguments.
What the seeding asked a deeper reading to check (NOTE-138 is that reading):
- It gives the “instance agent” as a precise unit for LLM psychology. That converges with Birch’s LIT-111 argument against a persisting interlocutor and contrasts with Roberts’s LIT-207 fictionalism, which makes it a natural hub for a nucleation cluster on LLM agency.
- The HHH+0 framework makes testable predictions about agentic misbehaviour driven by system prompts. A deeper reading should check which empirical results in §5 are cited as evidence against the simpler hypotheses.
- Compare the Inquiry version with the preprint, especially §§6–7.
Access when seeded: philpapers.org, philarchive.org and tandfonline.com returned 403. Read the preprint PDF (dated September 22, 2025, 28 pp.) from the Wayback Machine capture of 2026-03-21 of philpapers.org/archive/GOLWDC-2.pdf: abstract, §1 opening, all section heads, §3 (models vs instances), §5.6 (HHH+0), the opening of §6 (role play) and §7.4. §7.1–7.3 were read by heading only. The Inquiry version (online 4 Sep 2026 per Crossref, CC BY-NC-ND) was not reached, so differences from the preprint are unverified. published: is the preprint’s own date. The paper lists no keywords; those above are its own terms.