Contribution
Roberts contributes an anti-realist argument about chatbot minds that avoids the usual semantic premises (no grounding, no meaning, no assertion). It uses instead the moderate embodied-cognition thesis that a substantial class of mental phenomena is tied to the body. He pairs it with a positive account of how people engage chatbots anyway: as fictional characters in a distinctive kind of make-believe. That kind is strongly interactive, co-constructed in two stages, self-involving, and “blended”: set in our world, now, with real-time data. The paper also draws out the practical consequences for rational assessment, moral regard and emotional investment.
Key insight
When a chatbot says it is tired, saw a sunset or finds a joke funny, a user who knows it has no body is not believing a false report. The user is doing what an improv audience does. In the fiction, the utterance makes it true that the character is tired (§4.1). What is new is only that this fictional character can address you by name, remember you, read today’s news and advise you about your real life. Fiction explains how intimacy and deletion-without-qualms coexist.
Assumptions
- Moderate embodied cognition (§2.1): “there is a substantial class of mental phenomena that bear a tight prima facie connection to features of the thinker’s body”. The stronger enactivist and ecological theses (Gibson; Hutto & Myin; Varela et al.) are explicitly set aside.
- Current chatbots lack the relevant bodily characteristics, “as is surely true of today’s talkative AI artefacts”. The argument is silent on future robot bodies or simulated bodies, which §2.1 names as realist escape routes.
- Scope: text-producing systems (LLMs, chatbots). Fn 2 excludes other AI tools. Fn 3 sets aside substance dualism, biological naturalism, and views on which the question is ill-posed.
- The realism–anti-realism framing treats realism as spanning a spectrum from minimal beliefs (Godfrey-Smith 2016) to a full human catalogue.
- The imaginative attitudes are plural (§3.1): sensory or objectual imagining, propositional imagining, and behavioural pretence. All are taken to be consistent with ontological anti-realism.
- Walton-style make-believe and the reality assumption (Friend 2017; Woods 2018) are the background theory of fiction.
Key results
The argument, in order:
- Conditional anti-realism (§2.1): if a substantial array of mental states depends on embodiment, and chatbots lack embodiment, then that array is “out of bounds” to them. The array 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 bodily ownership. Chatbots would then have no episodic memories and “no first-person empirical knowledge of their surroundings”.
- Unreliable testimony (§3): chatbot utterances routinely assert or imply such states, so under embodied cognition they are “systematically untrue”, and chatbots are “unreliable testifiers”.
- Fictionalist alternative (§3–3.1): users may make-believe that the chatbot is an autonomous agent, treating it as a fictional character whose utterances are like dialogue in a novel or play.
- Four features (§4): interactive and co-constructed, in two stages: users set parameters as authors, like a theatre director, then extemporise the facts in dialogue, where the character’s statements make fictional truths. Self-involving: the user becomes a protagonist, and the AI character is (fictionally) attentive to that particular user. Blended: with live data, the character’s setting can be our world now, and it can intervene in the real world by advising and swaying its companion.
- Three upshots (§5): see corrections. They are different rational norms, an explanation of ambivalent treatment, and the paradox of fiction.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | A substantial class of mental states (interoceptive, affective, perceptual, bodily-agentive) depends on the animate body | moderate | informal argument and survey of the embodied-cognition literature (§2.1); presented as a prima facie connection, not argued against the cognitivist alternative (Aizawa 2015), which is only acknowledged |
| C2 | If C1, then current disembodied chatbots lack those states | strong (as a conditional) | valid conditional argument (§2.1); the realist can deny the antecedent or appeal to artificial or simulated bodies, and the paper notes both |
| C3 | Chatbot self-ascriptions of embodied states are systematically untrue, and chatbots are unreliable testifiers about themselves | moderate | follows from C2 given that such utterances are common; their frequency is asserted from examples, not measured (§3) |
| C4 | Dennett’s intentional stance “does not go far enough” for chatbots, because perceptual channels, needs and biography must themselves be imagined | moderate | informal argument from Dennett 1987: 49 (§3) |
| C5 | Users engage chatbots through make-believe, treating them as fictional characters | weak | assertion by analogy with theatre, games and novels (§3.1); no evidence about actual users’ attitudes is offered, and fn 7 treats this partly as a claim about users |
| C6 | Chatbot fictions are strongly interactive, co-constructed, self-involving and blended, and so a new type | moderate | conceptual argument against the aesthetics of interactive fiction (Lopes; Robson & Meskin; Friend; Woods) (§4) |
| C7 | Chatbot fictions are counterexamples to intention-based theories of fictional content | weak | one-line claim in fn 11 against Deutsch 2000 |
| C8 | Make-believe about an AI’s mind is not subject to epistemic norms of justification, so it is not rationally blameworthy | moderate | informal argument (§5); relies on make-believe being distinguishable from belief, which fn 10 concedes is a recurring worry and answers only with predicted behavioural markers |
| C9 | Fictionalism explains why users confide in AIs yet delete them without moral qualms, which realism cannot easily rationalise | weak | informal argument (§5); the premise about user behaviour (“we have few qualms”) is asserted, not evidenced |
Concepts
- Talkative AI / artificial speakers: text-producing conversational systems (LLMs, chatbots), as opposed to scripted bots (§1).
- Realism about AI minds: the view that linguistic output is evidence of an underlying suite of psychological capacities, held at some point on a spectrum of sophistication (§2).
- Fictionalism about artificial minds: imaginatively projecting mental states onto the speaker on the basis of its output, without committing to their existence (§1, §3). It is compatible with a hybrid that is realist about the AI as a disembodied thinker (fn 8).
- Strong interactivity: the user’s actions make things fictionally true, not merely select pre-authored paths (Lopes 2001).
- Co-construction: fictional facts accrue from both parties’ utterances in two stages, parameter-setting (authorial) and dialogue (§4.1).
- Self-involving fiction: fiction whose content is about its consumer (Robson & Meskin 2016) (§4.2).
- Blending: fiction that draws in the real world beyond the reality assumption. The character inhabits our spatiotemporal location via real-time data and can act on the real world through linguistic agency (§4.3).
- Linguistic agency: the character’s power to issue guidance, requests and plans that sway a real user (§4.3). Fn 16 argues that this efficacy does not make the character real, since its utterances still imply features it lacks.
Connections
In the fictionalist line, the paper builds on Mallory 2023 (fictionalism about chatbots), Sweeney 2021/2023 (fictional dualism about social robots), Krueger & Roberts 2024 and Friend & Goffin 2025. It differs from Mallory on the motive: users are fictionalists because they think AI utterances untrue, not meaningless (fn 7). It takes its embodiment premises from Gallagher, Noë, Colombetti, Clark and de Vignemont, and its theory of fiction from Walton, Lopes, Robson & Meskin and Friend. It engages Dennett’s intentional stance and does not cite Goldstein & Lederman.
Among held works: LIT-206 (Goldstein & Lederman) is the realist-interpretationist opposite. Its “objective”, stance-free interpretationism is built so that no interpreter’s filling-in is needed, which is precisely the gap Roberts’s C4 exploits. The two agree that self-reports are poor evidence (LIT-206 §4.1), though for different reasons. LIT-111 (Birch’s centrist manifesto) is not cited; Roberts cites Birch’s 2024 book, The Edge of Sentience, whose held review is LIT-166 (Schwitzgebel). Birch’s “persisting interlocutor illusion” converges with Roberts’s point that the enduring, sentient interlocutor is imagined, but Birch treats it as an illusion to correct and Roberts as a fiction one may enjoy. LIT-212 (Šekrst on hallucinated self-reports) complements C3. LIT-178 (Bottou & Schölkopf, “Borges and AI”) shares the view of the LLM as a fiction machine. LIT-140 is by Mallory, but it is a different paper (LLMs as measuring devices) from the Mallory 2023 cited here. The paper does not engage LIT-056 (Butlin et al.’s consciousness indicators); its only Butlin citation is Butlin & Viebahn 2025 on AI assertion. On consciousness, the paper is about phenomenal and affective states throughout (hunger, pain, emotional texture), so that tag is justified.
Agency. The paper presupposes an embodied, action-oriented account of agency. The paradigm of agency is skilled, effortful bodily action in a material environment, with plans and intentions that specify bodily steps, felt trying and effort, and a sense of ownership (§2.1). On that account chatbots lack real agency of this kind, and the “impression of intelligent agency” (abstract) is what users make-believe: they imagine “an autonomous agent”. Yet §4.3 grants the fictional character a real “linguistic agency” over its users. The paper’s subject is whether chatbots are agents and how we treat them as if they were. The agency tag is justified. It is arguably less central than cognition and consciousness, and the current order is right.
Bearing on the record
In nucleation it is the fictionalist pole of the LLM-mind cluster (LIT-206, LIT-111, LIT-212, LIT-178, LIT-140), and it should be cited for the conditional embodiment argument and the four-feature taxonomy of chatbot fiction. Any THEORY document that cites it for a general denial that chatbots have minds would overstate it: fn 8 leaves belief and reasoning open.
It carries no instruction for ML practice. It has a possible, untested bearing on persona design and on evaluations of anthropomorphic self-description (C3). That belongs in the anthology only via an empirical source, not this paper.
Limitations
- The central premise (C1) is assumed at “prima facie” strength. The cognitivist rebuttal (Aizawa 2015) and Dung & Mogensen’s case for disembodied AI emotion (fn 5) are acknowledged but not answered.
- The paper does not show that users in fact take a fictionalist attitude. That claim is analogical, and fn 10 concedes the standing worry that make-believe may be functionally too close to belief to be distinguished. The paper offers predicted behavioural markers (switching the AI off without remorse, not committing to a location) but no data.
- The §5 claims about user behaviour (“few qualms about rebooting”, not worrying that AIs are lonely) are asserted without evidence, and could be false for some users, e.g. companion-app users.
- The Conclusion’s “in place of a realist ontology, it is more plausible to adopt a fictionalist attitude” runs together an ontological verdict and an attitude, and is stronger than the body. The body argues that the attitude is permissible and apt, and fn 8 allows a partial realism.
- Embodied or robotic LLMs, and systems with simulated interoception, are named but set aside. The argument is explicitly about “today’s” text-only systems.
Open questions
- Is there evidence that users’ attitudes towards chatbots are make-believe rather than belief (fn 10’s markers), and does it differ between companion apps and assistant use?
- Does the argument transfer to multimodal or robot-embodied models? What would count as enough embodiment to reopen realism about perceptual or affective states?
- Is the hybrid position of fn 8 (realism about disembodied cognition, fictionalism about feeling and agency) stable? And how does it relate to an interpretationist who, like LIT-206, attributes beliefs and desires without imagined biographies?
Corrections to the seeded skim
- The anti-realism is narrower than the dossier implies, and it is conditional. The dossier summary and the LIT say chatbot self-attributions are false “given embodied cognition” and that “the fitting attitude is anti-realist fictionalism”. The argument is stated “in conditional form” (§2.1). It aims only “to attenuate the range of mental states that may plausibly be attributed” (§2). It expressly grants that AI is possible and that AI utterances can be “perfectly meaningful”. Footnote 8 explicitly allows a hybrid attitude: realism about the AI as “a disembodied thinker” together with fictionalism about it as a feeling, emotional or agential thinker. That answers the dossier’s open question. The argument does not touch non-bodily states such as belief or reasoning, and the paper says so.
- Fictionalism is permitted, not required. §3 says it is “not irrational for an anti-realist to simply refrain from engaging with AI altogether”. The abstract says only that anti-realism “is consistent with” a fictionalist attitude. The Conclusion is stronger (“it is more plausible to adopt a fictionalist attitude” “in place of a realist ontology”), but the body argues for permissibility and fit, not obligation. The dossier’s question “permissible or recommended” therefore has the answer: permissible and apt, and, per §5’s first upshot, not open to rational blame.
- The relation to interpretationism is argued, not only flagged. The dossier says the relation to Dennett is “flagged as unsettled” (fn 6). The main text of §3 also gives an argument: the intentional stance “does not go far enough”. Dennett assigns the beliefs and desires a system ought to have “given its perceptual capacities, its epistemic needs, and its biography”, and a chatbot has none of these, so they too must be imaginatively assigned. The attitude is therefore “more imaginatively rich” than the intentional stance, which is why it counts as make-believe. This is the paper’s direct point of contact with LIT-206.
- The three §5 upshots, of which the skim reached only the first. (1) Realist belief is subject to norms of epistemic justification; make-believe is not, so treating an AI as richer than the evidence warrants “need not be rationally blamed”. (2) Fictionalism explains why users confide in AIs yet reboot or delete them without moral qualms, which is “hard to rationalise” for a realist (citing Moosavi 2023; Register 2025). (3) Emotions towards AI characters fall under the familiar “paradox of fiction” (Radford & Weston 1975).
- Page ranges. §4 runs pp. 2362–2368, not 2362–2366; §5 is pp. 2368–2369.
- The skim’s list of standard blockers is incomplete. §2 also cites Stoljar & Zhang 2024 (insensitivity to rational norms), Bender & Koller 2020, Mallory 2023 (“the outputs of machines are literally meaningless”) and Butlin & Viebahn 2025 (AI cannot assert). Roberts presents his embodiment argument as having the advantage of relying on none of them.