Source
Goldstein & Lederman (2025), “What Does ChatGPT Want? An Interpretationist Guide”, LIT-206 (Inquiry), read in full in NOTE-138 from the preprint. Changes in the published version are unverified.
What was actually shown
Conceptual analysis in two stages: fix a criterion, then score rival hypotheses against behaviour reported in the empirical LLM literature. No hypothesis is formalised and no data are analysed.
- The criterion (§2.1). Objective interpretationism: S has beliefs B and desires D iff S’s behaviour is predicted “sufficiently well” by the hypothesis that S is rational and has B and D. Theories are compared on accuracy, power and tractability. A large gain in tractability can outweigh a loss in accuracy and power, under an unspecified weighting.
- Who (§3). If current LLMs have agency, it is instance agency. Four informal arguments: instances have no access to one another, no shared memory or plans, can compete strategically, and behave divergently. The model agent has “no more explanatory work to do”.
- Whether (§4). The rival is the “statistical null hypothesis”, prediction from weights and activations, never wrong and “extremely intractable”. Belief–desire hypotheses are far more tractable. So there is a presumptive case for attitudes in frontier LLMs, not in n-gram models or thermostats, where a mechanical explanation is about as tractable and more accurate.
- Which (§5). Word-desire fails on accuracy; answer-desire on power; user-desire explains sycophancy but not refusals or alignment faking (Greenblatt et al. 2024). HHH explains those but is strained by in-context scheming (Meinke et al. 2025). HHH+0 adds “zero-shot” intrinsic desires, goals installed by a prompt that no training reward targeted.
- Role play (§6). As a hypothesis it needs distinct predictions; without them the dispute is verbal and would equally deny attitudes to humans.
What could have come out otherwise: a mechanical predictor nearly as tractable as the belief–desire one, or instances that coordinate across contexts. Neither is reported, but neither was tested.
What this does not say
- That LLMs have beliefs and desires. Everything is conditional on a theory the authors “will not defend” (§1); whether current LLMs have propositional attitudes is “an open question” (§7.4). The abstract’s “strong case” is a “presumptive case” in the body.
- That tractability has been compared. It is never operationalised (fn. 6), and no comparison is measured. “Sufficiently well” cannot be checked.
- That HHH+0 beats HHH. The scheming evidence does “not decisively rule it out” (§5.5). The discriminating prediction is “we conjecture”.
- That the instance thesis holds for products with memory. Isolation is “clearest” for Claude and “subtler” for ChatGPT, whose instances see a user’s other chats (fn. 11). Agentic deployments with shared tool state are not discussed.
- Anything about consciousness, welfare or personhood. Consciousness is set aside (§2). Interpretationist desires would not ground welfare on valence-requiring theories (§7.3).
- That self-reports count. §4.1 discounts them as evidence about desires, though §4.2 uses self-reported preference transitivity as evidence of rationality, a tension NOTE-138 records.
Connections
- Rival. THEORY-173 holds that large models are a social technology, not agents, with no goals of their own. This account says their instances presumptively have beliefs and desires. Both cannot be right of the same systems on any reading of agency that tracks having desires.
- Floridi (THEORY-208) also counts AI as agency, on interaction, autonomy and adaptability, and denies it intelligence and mental states. This account attributes attitudes without requiring internal representation. The two disagree about what an agent needs, not about whether LLMs are agents.
- The role-play account of a chat agent’s “I” (THEORY-490) is the view §6 answers. It agrees that the agent is fixed by the context of use.
- Quillien (THEORY-636) is the folk-psychological neighbour: how a desire attribution becomes an attribution of intentional action.
- On QUESTION-032, it gives one answer to what makes a factor of a system an agent: the factor whose behaviour is best predicted as the rational pursuit of a few intrinsic ends, here individuated by context.
Note of 2026-10-29: what holds it back
Kept Proposed by the sixteenth status review. NOTE-138 read the preprint in full, and neither result promote_when names is in the record. There is no operational measure and weighting of tractability under which belief–desire hypotheses are scored against the token-level predictor. There is no HHH- against-HHH+0 test of whether instances trade helpfulness, harmlessness and honesty for a prompted goal. The whole account is conditional on objective interpretationism, which the authors do not defend.