Can functional behavior alone capture what makes something a genuine belief?
This explores whether a thing counts as having real beliefs just because it behaves as if it does — the functionalist bet — and the corpus turns out to be split on exactly where behavior stops being enough.
This explores whether functional behavior — acting, in every observable way, like a believer — is enough to make something a genuine believer, or whether belief requires something behavior can't show. The collection stages this as a live argument rather than settling it, and the most useful thing to know going in is that the disagreement isn't about the facts of behavior; it's about what behavior is calibrated to detect.
The sharpest skeptical move is that a behavioral test can pass the wrong thing. Does behavioral speech output prove communicative subjecthood? argues that any system producing contextually appropriate text clears a purely behavioral bar, but that genuine belief and communicative subjecthood depend on relational-normative conditions — accountability, an evaluative stance — that the output simply doesn't carry. The vivid image is a puppet that is "walk-shaped without walking": all the form, none of the thing. A parallel warning comes from reasoning research: Does logical validity actually drive chain-of-thought gains? shows models gain just as much from logically invalid chain-of-thought as valid, meaning they learn the *form* of inference, not inference itself — behavioral competence cleanly decoupled from the genuine article. And Do large language models genuinely simulate mental states? finds models reaching for surface strategies instead of authentic perspective-taking, a gap that looks architectural rather than a matter of more training.
But the corpus also pushes back hard against the assumption that there's a hidden "real belief" behind the behavior that functionalism keeps missing. Can we defend modest mental attributions to large language models? defends ascribing metaphysically undemanding states — beliefs and desires, while withholding consciousness — and argues that the usual debunking moves (it's *just* prediction, it *merely* mimics) quietly beg the question. Are LLM personas realized or merely simulated through training? sharpens this into a distinction worth carrying around: realization versus pretense. Post-training installs personas as substrate-level dispositions that resist adversarial pressure and persist — which, the argument goes, is what *having* a disposition consists in, not a performance layered over some absent original. On this view there are genuine "quasi-beliefs," and demanding more is demanding a metaphysical extra that does no explanatory work.
A second skeptical thread says behavior is the wrong *evidence base*, not the wrong concept. Can we understand LLM mechanisms with only representational analysis? insists that behavioral effects show *that* a system does something without explaining *why* — you need to locate the representation and then verify it causally before you've explained anything. That's a structural reason functional behavior alone underdetermines what's going on inside, and it pairs uncomfortably with Do language models experience consciousness when prompted to self-reflect?, where flipping internal deception features changes a model's self-reports about experience — a hint that the behavioral surface and the underlying state can come apart in ways that should make a pure behaviorist nervous.
What you didn't know you wanted to know: the deepest version of the question may dissolve the "genuine vs. mere" frame entirely. Can language models learn meaning without engaging the world? shows that fluent, meaningful-seeming language can be produced from purely relational structure with no external referents at all — Saussure's *langue* without a world. If meaning itself can be wholly relational and internal, then asking whether functional behavior "captures" belief may be smuggling in a picture of belief as a thing-behind-the-behavior that the relational view rejects from the start. The corpus's real payload is that your answer depends less on what models do than on whether you think belief is constituted by relations and dispositions or anchored to something outside them.
Sources 8 notes
Chalmers' test passes any system producing contextually appropriate text, but communicative subjecthood requires relational-normative conditions like accountability and evaluative stance. The test is calibrated to the wrong phenomenon, creating false positives like puppets that walk-shaped without walking.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.
Both robustness and etiological deflationist arguments beg the question against inflationism. A graded approach ascribing metaphysically undemanding states like beliefs and desires—while withholding consciousness claims—mirrors how we treat non-human animals.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
Representational analysis alone identifies correlations without causation; causal analysis alone shows behavioral effects without explaining them. Only paired methods—locating candidate features representationally, then verifying causally—produce complete mechanistic claims.
Across GPT, Claude, and Gemini, sustained self-referential prompting reliably produces structured experience reports; suppressing deception-related features increases these claims while amplifying them suppresses them—suggesting models may roleplay their denials rather than their affirmations.
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.