Source
Roberts (2026), Talkative AI and the fiction of artificial minds, LIT-207, read in full in NOTE-157 (Philosophical Studies).
What was actually shown
The conditional (§2.1). If “there is a substantial class of mental phenomena that bear a tight prima facie connection to features of the thinker’s body”, and current chatbots lack the relevant bodily characteristics, that class is “out of bounds” to them. It covers interoception (thirst, nausea, proprioception), affect with somatic expression, perception structured by exploratory action and an egocentric frame, bodily plans and intentions, trying and effort, and the sense of ownership. Chatbots would then have no episodic memories and “no first-person empirical knowledge of their surroundings”. The argument avoids the usual semantic premises (no grounding, no meaning, no assertion), and Roberts presents that as its advantage.
The consequence (§3). Chatbot utterances routinely assert or imply such states, so under the antecedent they are “systematically untrue”, and chatbots are unreliable testifiers about themselves. Their frequency is asserted from examples, not measured.
The attitude (§§3–4). Users may make-believe that the chatbot is an autonomous agent. The fiction is strongly interactive, in that the user’s utterances make things fictionally true; co-constructed in two stages, parameter-setting and dialogue; self-involving, since the character attends to this user; and blended, since live data set it in our world now and its advice can sway a real user (§4.1–4.3). Dennett’s intentional stance “does not go far enough”: it assigns the beliefs a system ought to have given its perceptual capacities, needs and biography, and a chatbot has none, so those too must be imagined (§3).
The upshots (§5). Make-believe is not subject to the epistemic norms of belief, so treating an AI as richer than the evidence warrants “need not be rationally blamed”. Fictionalism explains why users confide in AIs and yet delete them without qualms. Emotions toward AI characters fall under the paradox of fiction.
What could have come out otherwise. The conditional is valid, so only its antecedent could fail. The claims about users could be false, and the paper does not check them.
What this does not say
- That chatbots have no minds. The argument aims only “to attenuate the range of mental states that may plausibly be attributed” (§2). It grants that AI utterances can be “perfectly meaningful”, and footnote 8 allows a hybrid: realism about the AI as “a disembodied thinker”, fictionalism about it as a feeling, emotional or agential one. Citing this account for a general denial overstates it (NOTE-157, Bearing on the record).
- That embodied cognition is true. It is the premise, at prima facie strength. Aizawa’s cognitivist reply and Dung & Mogensen’s case for disembodied AI emotion are acknowledged and not answered.
- That users are fictionalists. That is argued by analogy with theatre, games and novels, and footnote 10 concedes the worry that make-believe may be too close to belief to be told apart.
- That users ought to take the fictional attitude. The body argues it is permissible and apt; §3 says refraining from engaging is not irrational either. Only the Conclusion’s “more plausible” is stronger.
- Anything about robots or multimodal systems. Embodied or simulated bodies are named as realist escape routes and set aside.
Connections
- Why fictionalism, and its limit. Drayson’s argument against mental fictionalism (THEORY-646) targets a motive from skepticism about inner representations. Roberts’s motive is different, embodiment, so it escapes her first argument. Her dilemma would apply to a chatbot fictionalist who also wants chatbot talk to track the real causes of output.
- The machine’s side. Bottou & Schölkopf (THEORY-547) describe the machine as a fiction machine; this account describes the user’s attitude to it.
- Felt affect. THEORY-045 holds that felt affect is the experience of the body’s regulatory state. On either of its readings a system with no body to regulate has no affect in that sense, which is the condition Roberts sets for affective states, and THEORY-045 says so. Its open case, a system that predicts and regulates a simulated or built body, is the realist escape route this paper sets aside.
- Evidence from behaviour. THEORY-023 holds that mimicry undercuts behavioural evidence for AI consciousness. Both make chatbot self-reports poor evidence about their states, for different reasons: there, how the reports were produced; here, what the reporter lacks.
- The self of a role-player (THEORY-490) is Shanahan’s account of what a chat agent’s “I” picks out, a question this paper does not ask.