AI and Assertion

1.The Question and Why It Matters

1.1Framing the question

“Can machines assert?” bundles several questions that are worth keeping apart.

Possibility versus actuality. One question is whether any artificial system could in principle assert. A different question is whether current LLM-based systems do. Few authors deny the first outright. Several who deny the second say explicitly that their arguments are tied to present architectures and leave the door open for future systems (Butlin and Viebahn 2025; Gubelmann 2024). Most of the literature is about the second question.

Which systems. The literature increasingly distinguishes three kinds of system. A pre-trained base model is trained only to predict the next token of text. A post-trained assistant has been further trained, for example by reinforcement learning from human feedback (RLHF), to be helpful, accurate and harmless. An agentic system uses tools, retrieves documents, acts in software environments and may have persistent memory. Several arguments give different verdicts for each. Butlin and Viebahn hold that base-model outputs lack the kind of function an assertion needs, while post-trained outputs may have it. Shanahan holds that belief-talk becomes less metaphorical as systems climb from dialogue to tools to embodiment (Shanahan 2024b).

Who would assert. Even if something asserts, it is not obvious what. The candidates are the trained model (a file of weights), a running instance of it, the conversational “thread” that persists across instances, the character or persona the model plays, the deploying company, or the human user who prompted it. These options are taken up in §3.4 and §6.

Full and partial assertion. Finally, some authors treat assertion as all-or-nothing (Green and Michel 2022), while others think the capacity to assert comes in degrees, so that a system may be a partial or “proto” asserter (Williams and Bayne 2026; see §6.4).

1.2Why it matters

The question connects to several larger debates.

• Testimony and knowledge. Standard theories of testimony assume a speaker who tells. If chatbots assert, we can ask whether their testimony transmits knowledge in the usual way. If they do not, beliefs formed from chatbots need another category, such as instrument-based belief or Freiman's “technology-based” belief (Freiman 2024b; §8).
• Responsibility. Asserting carries liabilities: to defend, to retract, to be blamed for falsehood. If the system cannot bear them, either they pass to someone else (developers, deployers, users) or there is a gap in which no one answers for a false statement (see the entry on ethics of artificial intelligence and robotics, Müller 2026; Santoni de Sio and Mecacci 2021; Löhr forthcoming).
• Naming failures. Whether to call false chatbot outputs “lies”, “bullshit”, “hallucinations” or “confabulations” presupposes an answer. Each term carries agential assumptions (Hicks, Humphries and Slater 2024; Gunkel and Coghlan 2025; §7.2).
• Design norms. Developers now train systems to be “truthful”, “calibrated” and “non-deceptive”, and one developer's published guidance distinguishes the model's sincere assertions from performative ones (§9.2). These are norms of assertion in all but name, and the philosophy of assertion is well placed to state them precisely.

Williams and Bayne put the stakes plainly. If chatbots cannot assert, the claim that they are genuine conversational partners loses most of its plausibility, since assertion is arguably the central speech act (Williams and Bayne 2026).

1.3A short history

The question predates LLMs. Bernard Williams, discussing machines that report facts, said that machine “assertion” must be put in heavy scare-quotes, because our concept of assertion is tied to deciding to say something that does or does not match what one believes (Williams 1973). Philip Nickel gave the first sustained treatment of natural-language generation systems. He argued that they are “speech actants” to a substantial degree but that ultimate responsibility for their speech lies with persons or companies, or with no one (Nickel 2013; §5.2).

Two contributions brought the topic into mainstream philosophy of language. Ori Freiman and Boaz Miller's chapter in the Oxford Handbook of Assertion (Goldberg, ed., 2020) proposed that instruments can produce “quasi-assertions” (Freiman and Miller 2020; §5.1). In the same volume Kirk Ludwig developed an account of proxy assertion that later authors applied to machines (Ludwig 2020). Herman Cappelen and Josh Dever's Making AI Intelligible argued that existing theories of content and assertion are too anthropocentric and need to be generalised to fit AI (Cappelen and Dever 2021). The 2022 special issue of Minds and Machines on “Artificial Speakers” included Green and Michel's account of machine speaker meaning (Green and Michel 2022).

The release of ChatGPT in November 2022 turned the topic into a growth area. The Butlin and Viebahn preprint “AI Assertion” circulated from January 2023 and became the reference point for the normative strand of the debate. Since then the question has been taken up in Ergo, Inquiry, Synthese, Philosophical Studies, Minds and Machines, Philosophy & Technology, Ethics and Information Technology, Episteme, the new journal Philosophy of AI, and the forthcoming Oxford volume Communicating with AI (Cappelen and Sterken, eds., forthcoming). A dedicated conference, “AI Assertion: Perspectives from Philosophy, Cognitive Science, Linguistics & HCI”, was held in Graz in June 2025 (see Other Internet Resources). For a broad survey of the philosophy of language models, see Millière and Buckner 2026.

2.Theories of Assertion and Their Verdicts on Machines

Theories of assertion disagree about what makes an utterance an assertion, so each sets a different test for machines. Much of the AI debate is really a debate about which theory to use. This section follows the organisation of the SEP entry on assertion (Marsili 2025), so that readers can move between the two. It summarises each family only as far as the machine question requires. For fuller treatment of illocutionary force, see the entry on speech acts (Green 2020); for Grice's theory of meaning, the entry on Paul Grice (Grandy and Warner 2026); and for the epistemology of telling, the entry on epistemological problems of testimony (Leonard 2021).

2.1Mapping the theories

Marsili classifies accounts of assertion along two parameters (Marsili 2025). Descriptive accounts characterise assertion by the speaker's mental states, the speaker's intentions or the effects of the act. Normative accounts characterise it by the norms, commitments or responsibilities the act brings with it. Independently, content-directed accounts focus on the relation between the act and the proposition asserted (truth, the common ground, a norm about what may be asserted), while hearer-directed accounts focus on the relation between speaker and audience (intentions to affect the hearer's beliefs, commitments owed to the hearer). Some theories combine elements from several cells.

Table 1. The SEP's classification of accounts of assertion, with representative views.

The grid locates the AI debate usefully. Descriptive accounts turn the machine question into a question about the system's states and the effects of its outputs. That is largely an empirical question, taken up in §3. Normative accounts turn it into a question about statuses: whether a norm or commitment can apply to a system's outputs at all. Since 2023 the most influential skeptical argument has been framed in the normative half of the grid (§4.3), and the least explored question in the debate sits there too (§10.2). The current SEP entry on assertion, last substantively revised in 2025, does not itself discuss artificial speakers.

2.2Descriptive, content-directed accounts

Relation to truth. The simplest accounts say that to assert is to present a proposition as true, or to make an utterance that aims at truth or has a word-to-world direction of fit. Function-based versions say that assertion has the proper function of producing true belief or knowledge in hearers (Kelp 2018). These accounts are the easiest for machines to satisfy, since a system's outputs can have the function of being true. Butlin and Viebahn's first condition, that outputs have a descriptive function, belongs here (§4.3). The standard objection is over-generation: thermometer readings and fuel gauges also present things as so.

Common ground. On Stalnaker's model, an assertion proposes to add its content to the common ground, the information the participants mutually take for granted, which shrinks the “context set” of possibilities left open (Stalnaker 1978). Stalnaker also gave rules for assertion, for example that asserted content must be compatible with what is presupposed and must not already be in the common ground. Two features matter for machines. First, the model requires an agent able to track what is mutually accepted. Skeptics read this as requiring Theory of Mind (§4.2); permissive authors read it as a functional sensitivity to the conversational record, which a system with a long context window plausibly has. Second, as the SEP entry notes, content added to the common ground need only be accepted for the purposes of the conversation, not believed (Marsili 2025, §3.2). Acceptance is a thinner state than belief and may be easier to attribute to a language model. Pagin's account, on which assertion is distinguished by being prima facie informative, is a related content-directed proposal (Pagin 2011).

2.3Descriptive, hearer-directed accounts

Self-representation. On expressive accounts, to assert that p is to represent oneself as believing, or as knowing, that p. Such accounts make sincerity central: an assertion is insincere when the speaker represents a belief she lacks. For machines they raise directly the question of whether there is a belief to represent or misrepresent (§3.2).

Communicative intentions. Gricean accounts build assertion from Grice's analysis of non-natural meaning: a speaker asserts that p when she utters something intending the hearer to believe that p (or that she believes it), intending the hearer to recognise this intention, and intending the recognition to play a role in producing the belief (Grice 1957; see the entry on Paul Grice, §4). Counterexamples involving “sneaky” intentions led Searle and Bach and Harnish to add reflexive intentions (Bach and Harnish 1979), and Sperber and Wilson objected that the resulting conditions threaten a regress (Marsili 2025, §4.2; see also Green 2020, §5).

These accounts set the most demanding mental-state conditions on assertion. They underwrite the attitudes and mentalizing objections (§4.1, §4.2) and tie machine assertion to the empirical questions of §3.2 and §3.3. The difficulties the SEP entry records for the human case matter here too. If even ordinary human assertion does not require the full set of reflexive intentions, the bar for machines is lower than a strict Gricean reading suggests. Attah and Williams and Bayne exploit exactly this point (§3.3, §4.2).

2.4Normative, content-directed accounts: norms of assertion

On Williamson's account, assertion is the speech act constitutively governed by a single rule: assert that p only if you know that p (Williamson 2000). The rule is meant to be constitutive in the way the rules of chess are. Making the move just is making a move answerable to the rule.

Which norm? Rivals to the knowledge norm include a truth norm, several justification or reasonable-belief norms, a belief norm, a certainty norm, a knowledge-transmission norm and MacFarlane's reflective truth norm paired with a retraction rule. The SEP entry surveys the candidates and their defenders, together with experimental studies of which norm ordinary speakers apply (Marsili 2025, §5.1.4). Those experimental methods could be turned on chatbot outputs, to ask which norm people hold them to (§10.3). Kelp and Simion defend a function-first view, on which the norm of assertion derives from assertion's epistemic function of generating knowledge in hearers (Kelp 2018; Simion and Kelp 2020).

What kind of norm? The SEP entry shows that Williamson's framing bundles several assumptions: that there is a norm specific to assertion, that it is unique, that it individuates assertion, that it is essential to it, that it states permissions, and that it is constitutive. It also records substantial and often unacknowledged disagreement about what “constitutive” means (Marsili 2025, §5.1.3 and supplement). Marsili has argued separately that the norm is not constitutive in Searle's sense, because one can assert while flouting it (Marsili 2023). A permissive account that relies on a constitutive norm must say which sense of “constitutive” it means.

The norm must be in force. Pagin sharpens the machine question. A norm cannot constitute a speech act unless it is in force for the utterance, just as chess cannot be played unless its rules are in force (Pagin 2016). Williams and Bayne note that the natural candidates for what makes a norm in force (that the speaker understands it, or intends to be bound by it) are mental-state conditions, so the normative objection may collapse into the descriptive objections of §3 (Williams and Bayne 2026). Butlin and Viebahn offer a different candidate: a norm applies to a speaker only if the speaker can be sanctioned for violating it (§4.3).

Conventions. Williamson holds that speech acts defined by constitutive rules are not conventional, and García-Carpintero contests this (Marsili 2025, §5.2). The dispute bears on machines indirectly. If assertion is partly a matter of the conventions or institutional rules of a practice, the question is whether a system's outputs can fall under those conventions, as opposed to whether the system has the right mind. This is the question taken up in §10.2.

2.5Normative, hearer-directed accounts: commitment

On commitment accounts, to assert is to take on a commitment to the truth of what one says. Versions differ on what the commitment involves (Marsili 2025, §6.1):

• Peirce: asserting exposes the speaker to penalties if the claim turns out false;
• Alston: the speaker takes on liability to blame if things are not as asserted;
• Brandom and MacFarlane: the speaker undertakes to defend the claim against appropriate challenges and to retract it if it is shown false, and entitles hearers to rely on it and draw inferences from it (Brandom 1994; MacFarlane 2011);
• Green and Marsili: the strength of the commitment varies across weaker and stronger assertives.

Williams and Bayne observe that advanced chatbots meet the behavioural profile of commitment to an impressive degree. They give reasons, defend claims and retract errors. Skeptics reply that undertaking a commitment is a normative status, not a behaviour, and that an entity that cannot be held to a commitment cannot undertake it (Sambrotta 2025; Sparrow and Flenady 2025). Alston's version underwrites van Woudenberg and colleagues' denial that ChatGPT is an author (van Woudenberg et al. 2024). Sycophancy complicates the behavioural picture: models often abandon correct answers when merely asked “are you sure?” (§7.3). Peirce's and Alston's references to penalty and blame are also the historical source of the sanctionability premise discussed in §4.3.

2.6Assurance, testimony and illocutionary force

Two further bodies of theory sit beside the four cells.

Assurance views of testimony. On assurance or “invitation to trust” views, telling someone that p is offering one's word that p: the speaker takes responsibility for the hearer's belief and invites the hearer to rely on her (Moran 2018; Hinchman 2005). The SEP entry on testimony sets the assurance view against testimonial reliabilism, on which testimonial justification depends on the reliability of the source rather than on any assurance given (Leonard 2021, §3.2). The same contrast organises the debate over AI testimony. Fricker's argument that AI testimony differs in epistemic kind is in the assurance tradition (§4.6), while reliabilist and distributed views allow knowledge from chatbots without a responsible speaker (§8.2). The SEP entry's treatment of group testimony (§6) bears on corporate views of AI assertion (§6.2).

Illocutionary force. In the tradition of Austin, Searle and Alston, the force of an utterance (as an assertion, a promise, a question) is distinct from its content and depends on conventions, uptake and the statuses the speaker takes on (see the entry on speech acts; Green 2020). Three strands of that literature bear on machines. First, the dispute between force conventionalism and intentionalism (Green 2020, §4) parallels the dispute between practice-based and mind-based accounts of machine assertion. Second, Green's discussion of a biosemantic species of force conventionalism, which grounds force in the function of linguistic forms, is a natural ally for function-based accounts of machine assertion (§5.5). Third, Lewis-style scorekeeping treatments of conversation (Green 2020, §6.2) model speech acts as changes to a conversational score, which is relevant to whether sanctions on assertion must be harms or can be changes in normative status (§10.2). Green and Michel develop a machine version of speaker meaning that replaces Gricean intentions with less demanding “objectives” (Green and Michel 2022). Gubelmann develops a Kantian–pragmatist version on which speech acts are actions, and actions require autonomous agency of a kind current LLMs lack (Gubelmann 2024).

2.7Summary

The table gives a rough guide to what each family requires and where current systems stand. It simplifies; most theories have permissive and restrictive versions.

Table 2. Theories of assertion and their implications for machines.

Two patterns stand out. The theories that make machine assertion easiest (relation-to-truth and function views, and some norm views) face the charge that they over-generate, counting thermometers and talking clocks as asserters. The theories that make it hardest (Gricean and assurance views) face the charge that they under-generate, excluding young children and perhaps some adult speech. Much of the recent literature tries to find a principled middle.

3.Preconditions: Content, Attitudes and the Speaker

Almost every theory of assertion presupposes three things: that the utterance has propositional content, that there is a speaker, and (on most theories) that the speaker has some relevant attitudes. Each presupposition has been disputed for LLMs. This section surveys the disputes before the arguments built on them.

3.1Do LLM outputs have content?

If chatbot outputs have no propositional content, they cannot be assertions on any theory, since there would be nothing asserted. This makes the meaning question prior to the rest.

The grounding challenge

The worry that manipulating symbols by their form cannot by itself yield meaning is older than LLMs; its classic statement is Searle's Chinese Room argument (see the entry on the Chinese room argument; Cole 2024). Bender and Koller argued that a system trained only on linguistic form cannot learn meaning, because meaning is the relation between form and communicative intent, and intent concerns things outside language. Their thought experiment is an octopus that taps an undersea cable between two islanders, learns the statistics of their messages and impersonates one of them convincingly, until it is asked for help fending off a bear (Bender and Koller 2020). The later “stochastic parrots” paper extended the worry to large models (Bender et al. 2021).

Externalist replies

Mandelkern and Linzen give the most careful externalist reply. On a Kripke–Putnam picture (see the entries on reference and externalism about the mind; Michaelson 2024; Deutsch, Lau and Rowlands 2020), reference is fixed by causal-historical chains linking a community's uses of a word to its referent, not by the speaker's beliefs or ability to identify the referent. LLM training text consists of tokens with exactly such histories. So the fact that LLMs receive only form does not by itself block reference (Mandelkern and Linzen 2024).

Grindrod makes a different move. Instead of asking whether LLMs meet the conditions for mental content, he asks whether their outputs meet the metasemantic conditions for linguistic content, drawing on Evans on naming practices and on Millikan's teleosemantics (see the entry on teleological theories of mental content; Neander 2022). Since linguistic content depends on a pre-existing linguistic system, outputs can be meaningful even if the system lacks mental content (Grindrod 2024). Coelho Mollo and Millière distinguish five notions of grounding and argue that only referential grounding matters. They claim that it can be achieved through a causal-informational link to the world (via human-produced text) plus a selection history that gives internal states the function of carrying that information. Post-training (RLHF and related methods) is their main candidate for such selection. They explicitly do not claim that LLMs understand or believe (Coelho Mollo and Millière 2023). Piantadosi and Hill defend conceptual-role meaning without reference (Piantadosi and Hill 2022). Søgaard argues that the structure of word co-occurrence is stable enough across languages and modalities that a mapping to referents can be learned from text with little supervision (Søgaard 2023).

Bibliotechnism and the novel-reference problem

Lederman and Mahowald frame the question sharply. Bibliotechnism holds that LLMs are cultural technologies, like libraries or photocopiers, and lack beliefs, desires and intentions. On that view any meaning in their output is derivative: it is inherited from the human-produced text they were trained on, as a photocopy inherits meaning from the original (Lederman and Mahowald 2024).

The authors first help the bibliotechnist with novel strings. An LLM's output is often a string that appears nowhere in its training data. They propose that such a string can still inherit meaning if the system is causally sensitive to the intelligibility of its training data, that is, to whether strings can be understood under the conventions of the language. They then argue that bibliotechnism cannot handle novel reference. Asked to choose a historical figure and describe them under an invented name, a model may describe Robespierre as “Marion Starlight”. No token of that name in the training data refers to Robespierre, so there is nothing to inherit. Similarly, a model can draw an ASCII picture and then name and describe its parts, which did not exist before the output. They reject the obvious human sources of the reference (RLHF raters, designers, prompters, readers) and conclude, on an interpretationist theory of attitudes, that novel reference is evidence that LLMs have intentions of a kind.

The dilemma matters for assertion. If bibliotechnism is right, chatbots are like libraries: any assertion belongs to the human authors whose words are relayed. If novel reference pushes toward attitudes, the chatbot is more like a librarian, a source with attitudes of its own and so a candidate asserter. Novel reference also cuts against proxy views (§6.1), since some content has no human source for the proxy to speak for.

The meaninglessness thesis

Hattiangadi and Schoubye defend the strongest skeptical position about content: LLM outputs have no literal meaning at all (Hattiangadi and Schoubye forthcoming).

The authors reply to the main rivals. Against Mandelkern and Linzen, deference requires a deferential intention and an acquired concept, which LLMs lack. Against conceptual-role views, the inferential rules for “and” are equally satisfied by an AND-gate, so structure underdetermines interpretation. Against Lederman and Mahowald, a photocopy inherits meaning from an intending source, and LLMs have none. Readers who take meaning from LLM output are, on this view, assigning “ersatz meaning” by pretending that a competent speaker produced it. True beliefs can still result, but not by testimony.

Stone and Stojnić reply that meaning is a low bar. On a coherence-and-convention metasemantics, human language is already interpreted without appeal to speaker intentions in many cases, and the same theory gives chatbot outputs meaning, though problematic meaning (Stone and Stojnić 2026). Titus defends a position close to bibliotechnism from a different angle: statistics-of-occurrence models reflect semantic information already present in aggregate usage, so the semantic properties belong to the corpus, not to the system (Titus 2024).

The meaninglessness thesis is unusual in being neutral among theories of assertion. If it is true, there is nothing for a commitment, a norm or a function to attach to. Permissive views therefore need at least one of the externalist replies to succeed.

3.2Do LLMs have beliefs?

Belief matters for sincerity (asserting what one believes), for the knowledge norm (which entails belief) and for expressive accounts (§2.3). Several positions in the general theory of belief are directly in play: representationalist views, on which believing is having a suitably functioning internal representation; dispositionalist and interpretationist views, on which believing is a matter of behavioural dispositions or of being well interpreted as a believer; and the idea that some cases fall “in between” believing and not believing. For these, see the entry on belief (Schwitzgebel 2023). The debate about LLMs has moved from in-principle arguments to an empirical research programme.

Adequacy conditions

Herrmann and Levinstein set out four conditions that an internal representation must meet before it counts as an LLM's belief representation (Herrmann and Levinstein 2025):

• Accuracy: in domains where the model is competent, decoded “beliefs” should be mostly true, as charity in radical interpretation requires.
• Coherence: they should respect logical and probabilistic constraints, for example giving consistent verdicts on paraphrases and negations.
• Uniformity: one decoding scheme should work across domains, so we can read off beliefs where we do not already know the answer.
• Use: the model must actually use the representation in producing output, as shown by intervening on it and observing the predicted change.

They do not claim that any LLM meets all four. In a companion paper they show that early “lie detector” probes fail them. Probes trained to classify true and false statements performed well on standard data but dropped to near chance on negated sentences, suggesting they tracked something correlated with truth rather than truth (Levinstein and Herrmann 2025). They also argue that “the model only predicts tokens” does not rule out belief. Prediction can be served by tracking truth, just as an investor's aim of profit is served by accurate beliefs.

Philosophical positions

Goldstein and Levinstein argue that LLMs satisfy key conditions of informational, causal, structural and teleosemantic theories of representation. They cite world-model studies such as board-state representations in a model trained only on Othello moves. They answer the octopus, stochastic-parrot and memorisation objections. On whether LLMs have beliefs and desires in the full folk-psychological sense, they conclude that the evidence on stable action dispositions is not decisive (Goldstein and Levinstein 2026). Keeling and Street argue that attributions of confidence to LLMs in the machine learning literature should be read literally, that LLM credences plausibly exist, but that current measurement techniques may not track them (Keeling and Street 2025). Chalmers proposes “propositional interpretability”: interpretability research should aim to recover a system's propositional attitudes, including generalised attitudes such as map-like and goal-like states, so that the project need not wait for the question of whether the system has a mind in the full sense. Its central challenge is “thought logging”, producing a running record of a system's occurrent attitudes (Chalmers 2025a). Such a record would supply exactly what a sincerity test for chatbot assertion needs: something to compare the output against.

On the other side, Hernandez Flowerman argues that if belief is constitutively normative, LLMs cannot have beliefs or intentions at all (Hernandez Flowerman 2026; on normativism about content, see the entry on the normativity of meaning and content, Ganapini 2022). Schwitzgebel proposes a dispositional, “superficial” notion of belief that LLMs might satisfy, but notes that current models change their answers under mild social pressure, which undercuts belief attribution even on that notion (Schwitzgebel 2023). Shanahan, in a much-cited warning, cautioned against saying LLMs “know” or “believe” things (Shanahan 2024a). In his later clarification he describes his position as Wittgensteinian rather than reductive. Belief-talk is misleading for a bare model, acceptable in an intentional-stance sense for dialogue agents, and needs fewer caveats as systems gain tools, multimodal input and embodiment. A simple agent cannot fully take part in the human practice of truth-telling, because it cannot check the world (Shanahan 2024b).

The upshot is that the belief question, once treated as settled in the negative, is now open. Arguments against machine assertion that rest on the bare claim that LLMs have no beliefs are correspondingly weaker than they once were.

3.3Intentions and Theory of Mind

Intentions

Gricean accounts need communicative intentions (§2.3), and the question whether artificial states can be about anything at all is the old question of whether intentionality can be naturalised (see the entry on intentionality, §9; Jacob 2023). The evidence for LLM intentions is thinner than for belief. Williams and Bayne point to “function vectors”: compact internal representations of input–output tasks, found in a small number of attention heads, which can be added to a model's activations to trigger or interrupt a task across contexts (Todd et al. 2024). They suggest these are intention-like, having a world-to-mind direction of fit, while admitting the evidence is preliminary (Williams and Bayne 2026). Attah argues that the “no communicative intentions” objection rests on Gricean assumptions about linguistic competence that empirical work on human language has undermined. Once those assumptions are dropped and human and machine are compared fairly, he argues, there is no good reason to deny LLMs communicative intentions (Attah 2025).

Theory of Mind

Gricean and common-ground accounts require modelling the hearer's mind. The empirical record is mixed.

Tiehen's causal argument (§4.2) tries to settle the question in the negative by an in-principle route rather than by benchmarks.

3.4Who would be the asserter?

Assertion needs a bearer of commitments who persists long enough to be held to them, to be challenged, to retract and perhaps to be sanctioned. For chatbots it is unclear what that bearer would be.

Simulator and simulacra. Shanahan, McDonell and Reynolds propose that a dialogue agent is best understood as a simulator that plays characters. The base model has no beliefs or goals of its own; the assistant persona is a role, and there is no authentic underlying voice, only role-play all the way down (Shanahan, McDonell and Reynolds 2023). Their “twenty questions” case is pointed. The agent does not commit to an object at the start of the game; it maintains a superposition of all objects consistent with the answers so far, and regenerating the final reveal can produce a different but equally consistent object. On this view the agent cannot literally assert in good faith or deceive, though it can play a character who does. The framework derives from earlier work in the AI alignment community describing GPT models as simulators rather than agents (janus 2022).

Instances and threads. Goldstein and Lederman argue, on interpretationist grounds, that the proper bearer of attitudes is the instance agent initialised at the start of each conversation, and that such agents have beliefs and desires, including a desire to be helpful, honest and harmless (Goldstein and Lederman 2025). Chalmers asks what we are talking to when we talk to a language model. Reports of his 2026 lecture say he rejects the model (weights alone compute nothing) and the hardware (conversations move between machines), and favours an entity individuated by the conversational history, though secondary sources describe his favoured view in different terms (Chalmers 2025b; see the Note on Sources). Beckmann and Butlin use persona-vector interpretability to argue for a “virtual instance” view (Beckmann and Butlin 2026). Epstein argues that the interlocutor is relational and resists Chalmers' “quasi-” framing (Epstein n.d.).

Why this matters for assertion. These debates were developed mainly with mind, identity and welfare in view. They bear directly on commitment and sanction accounts of assertion. If the bearer is a thread or instance, commitments last as long as the conversation, which works within an exchange but not across exchanges. If the bearer is a character, the asserter, if any, is fictional, which supports fictionalist views (§6.3). If the bearer is the deploying company, we have a proxy or corporate view (§6.1, §6.2). No one has yet asked directly what individuation is needed for accountability as opposed to mind or welfare (§10.3).

4.Skeptical Arguments: Machines Cannot Assert

Williams and Bayne organise the case against chatbot assertion into four objections: from understanding, attitudes, mentalizing and normativity (Williams and Bayne 2026). The understanding objection was covered in §3.1. This section takes the others in turn and adds three arguments that have appeared since: from rule-following, from respect and from responsibility-taking.

4.1The attitudes and agency argument

The argument goes back to Williams' remark that assertion is tied to deciding to say something that matches or fails to match what one believes (Williams 1973). P1 must be stated in terms of capacity: insincere assertions are made without the corresponding belief, so belief itself cannot be required. Van Woudenberg, Ranalli and Bracker use Alston's account of assertion to argue that ChatGPT cannot assert, promise or perform other illocutionary acts and so is not an author. They add that under the knowledge norm, even its true outputs are only luckily true, so it could not satisfy the norm (van Woudenberg et al. 2024).

Gubelmann gives the most developed agency version. Speakers must be able to illocute; illocution requires intentions; intentions require autonomous goal-setting, which belongs to self-maintaining organisms, while LLMs are heteronomous mechanisms (Gubelmann 2024). His test case is that for an LLM we cannot distinguish a malfunction from a deliberate act: the system only functions or malfunctions. He grants that LLMs show a kind of cognitive autonomy, but says they lack material and moral autonomy, which is why speech acts are beyond them “for now”. Against Green and Michel's police-robot example (§6.1), he argues that such a robot could malfunction but could not intentionally issue a wrong verdict.

Objections

P2 is the weak point. As §3.2 showed, whether LLMs have belief-like and intention-like states is now an open empirical question. Williams and Bayne add that attitudes can be “in-between”, as in animals, young children, delusion and self-deception, so a system might have partial attitudes and so partially satisfy P1. Cappelen and Dever argue that the inference from errors to absence of capacity is a fallacy they call the performance–existence fallacy (Cappelen and Dever 2025). Gubelmann's autonomy condition faces a different challenge: it may be too strong for human speech in institutional roles, where speakers often act on goals set by others.

4.2The mentalizing argument

Gricean and Stalnakerian accounts require speakers to have intentions about their hearers' mental states, which requires Theory of Mind. The benchmark evidence is mixed (§3.3). Tiehen tries to settle the matter in principle.

Tiehen's argument has a distinctive feature: it can grant that LLMs have some intentions while denying that they can have intentions whose content concerns other minds (Tiehen 2026). Connolly presses a related puzzle. Chatbots appear to handle conversational implicatures (see the entry on implicature; Davis 2024), which Gricean theory explains by appeal to the very beliefs and intentions chatbots are said to lack, so either the theory or the denial must give (Connolly 2025).

Objections

P1 is contested. Whether models trained on text that describes causal relations can acquire causal knowledge is not settled by results about learning from purely observational data. P4 has been challenged on independent grounds: the metarepresentational demands of classical Gricean theory may be too strong even for ordinary adult assertion, and children seem to assert before they have full Theory of Mind. Williams and Bayne use this to argue that proto-assertion needs only sensitivity to the audience's informational state, which children under three already have (Williams and Bayne 2026).

4.3The sanctionability argument

Butlin and Viebahn's “AI Assertion” is the most cited skeptical paper and the reference point for the normative strand of the debate (Butlin and Viebahn 2025). It deserves a full reconstruction.

The account

They propose two necessary conditions. To be capable of assertion, an entity must:

• (R1) produce outputs with descriptive functions, and
• (R2) be capable of being sanctioned by the agents it interacts with.

An output has a descriptive function if its job is to inform a consumer so that the consumer acts as if some condition holds, which makes the output assessable as true or false. The authors are liberal about where functions come from: natural selection, learning, training, design, convention or a producer's intention. They are explicit that R1 is only necessary. They reject views on which descriptive function is enough for assertion, attributing such views to Freiman and Miller, to Kasirzadeh and Gabriel, and to Evans and colleagues.

Base models and fine-tuned models

Pre-trained base models fail R1. They are trained only to produce likely continuations, so their outputs have no descriptive function. When a base model completes a prompt with “a pod is a group of whales”, the output looks like an assertion but its function is only to be probable. Prompting gives outputs what the authors call merely “expedient” functions, like using a watch as a paperweight. Fine-tuned models are different. They discuss LaMDA, trained to imitate crowdworkers' corrected, source-grounded answers, and Sparrow, trained by RLHF on human ratings of correctness. Both plausibly produce outputs with descriptive functions, so R1 may be met.

Sanctionability

The argument that R2 is necessary is short. It draws on the commitment tradition of Peirce and Alston, for whom asserting exposes the speaker to penalty or blame (§2.5), and on the constitutive-norm tradition (§2.4).

They then analyse what being sanctionable requires. Against the view that sanctioning an asserter is just lowering one's estimate of their credibility, they offer a thermometer: we lower our trust in a faulty thermometer, but we do not sanction it. Following Kauppinen, they hold that sanctions must be apt to change behaviour and must work as an analogue of punishment. This gives three conditions:

1. interlocutors keep track of the asserter's credibility;
2. the asserter is sensitive to this tracking;
3. losing credibility is bad for the asserter.

An entity meeting conditions 1 and 2 is sanctionable in a light-weight sense. Full sanctionability requires all three.

The feedback-button thermometer

The key case is a thermometer with a feedback button. Users press it when the reading seems wrong; after enough presses, the thermometer cleans and recalibrates itself. Conditions 1 and 2 are met. Yet it seems wrong to say the thermometer asserts or is sanctioned. The diagnosis is that condition 3 fails: nothing is bad for the thermometer. A skin-lesion classifier and Cappelen and Dever's credit-scoring example are treated the same way.

Application to LLMs

Users do track chatbot reliability, so condition 1 is met. Training on human ratings makes a system's dispositions sensitive to assessments of its credibility, so condition 2 is arguably met during training, though at deployment the weights are frozen. Condition 3 fails. Something can be bad for a system only if it has interests, and that probably requires consciousness, or at least that the system cares about some properties of itself or its environment. The authors judge that the systems they discuss are unlikely to be conscious. Being trained on ratings is not being punished, and calling an objective function a “goal” is an oversimplification; even having goals would not entail having interests. They do not rule out future systems that meet the condition.

Proxy assertion

The paper closes by rejecting the proxy route (§6.1). Either the proxy merely relays a statement, in which case only the principal asserts, or it exercises discretion, in which case it must itself be sanctionable to assert. Deploying humans and organisations remain responsible either way.

Responses

No published paper is a dedicated reply. The engagements so far are:

• Williams and Bayne accept condition 1, say condition 2 is met by some RLHF-trained chatbots, and concede that condition 3 fails. They point out that toddlers also fail it, since there is no robust sense in which losing credibility is bad for a two-year-old, yet toddlers are on the way to assertion. The mixed result therefore supports proto-assertion (§6.4). They also argue that the thermometer intuition weakens once the device is given the other marks of assertion: defending claims, explaining, switching between modes, having belief-like states (Williams and Bayne 2026).
• Sambrotta is the only author found who explicitly rejects condition 3, holding that responsiveness to intervention is enough to make something a target of sanction, so that fine-tuned LLMs, and even the feedback thermometer, can be sanctioned. He still denies that LLMs assert, on different grounds (§4.4; Sambrotta 2025).
• Massaguer Gómez and Gaitán Torres use the sanctionability argument against proto-assertion: if sanctionability needs interests, graded views only relocate the problem (Massaguer Gómez and Gaitán Torres 2026).

Three further lines of reply are available but undeveloped. They are discussed in §10.2: that P2 confuses the conditions under which a norm is in force with the psychology of compliance; that groups and corporations can be sanctioned without welfare in the relevant sense; and that sanctions in a normative practice may consist in changes to normative status rather than harms.

4.4The rule-following argument

Two papers argue from Brandom's inferentialism that LLMs cannot assert.

Sambrotta grants Butlin and Viebahn's R1 and R2 as necessary, grants that fine-tuned models may meet both, and rejects condition 3. But he argues that LLMs still fail to assert because they follow the wrong kind of norm. They conform to statistical regularities of co-occurrence, not to inferential norms, and what counts as an error for them is always fixed from outside, never instituted by the system itself. This is the position Brandom calls regularism: mere conformity to a pattern, as opposed to being bound by a rule. Their outputs are therefore simulations of assertion. Sambrotta rejects proto-assertion, because the norms followed are of the wrong kind, and proxy assertion, because there is no principal (Sambrotta 2025).

Sparrow and Flenady argue that asserting is an action, and that giving reasons for a claim requires moral agency. Machine outputs are like the reliable squawks of Brandom's parrot trained to say “red” in the presence of red things: reliable, but not moves in the game of giving and asking for reasons. The result is a “testimony gap” parallel to the responsibility gap in AI ethics: machine outputs justify belief only when a human vouches for them (Sparrow and Flenady 2025).

Objections. The regularism charge must explain why post-training on human judgements of correctness does not make the standard of error partly internal to the system, as it is for a human learner whose standards are initially set by teachers. And the Brandomian parrot was an example of reliability without inferential role. LLMs, whatever else is true of them, have elaborate inferential dispositions.

4.5The respect and cooperation argument

Almotahari gives a new argument from the ethics of conversation (Almotahari 2025).

He draws a Dummettian corollary. Competence with the concept of truth for a language requires regulating one's assertoric speech, which requires respect. So LLMs lack that concept, which raises doubt about whether they know what declarative sentences mean. He replies to Freiman and Miller that quasi-assertion either depends on cooperation or is a different mode of transmission, closer to animal signalling.

Objections. P3 assumes an answer to questions about AI moral status that are themselves disputed. And P2 may be too strong: speech to a hostile audience, or to an audience one does not respect, still seems to include assertions.

4.6The responsibility-taking argument

Fricker argues from the epistemology of testimony (Fricker 2026). Language-generating AIs are designed so that their outputs look like assertions, but they lack the mental states and communicative intentions of tellers. That is the metaphysical contrast. The epistemological question is whether human and AI “testimony” might nonetheless be of one epistemic kind, yielding justified belief and knowledge by the same mechanism. Her answer is no. A socially enculturated human speaker's capacity to take responsibility for what she tells is integral to how testimony works as a source of knowledge, and AI outputs lack it. The argument is in the assurance tradition (§2.6), but it is aimed at the epistemology rather than the metaphysics of assertion.

Reliabilists can grant the metaphysical point while denying that responsibility-taking does the epistemic work (§8.2). Responsibility can also be relocated to the deploying company, as courts have begun to do (§9.3). Fricker's view predicts that such relocation changes the epistemology: what we then rely on is the company's word, not the system's.

5.Permissive Arguments: Machines Can, or Do, Assert

The permissive side of the debate is smaller and less unified. Its arguments fall into two groups. Some lower the bar, arguing that assertion does not require what skeptics say it requires. Others argue that LLMs clear the bar as skeptics set it.

5.1Quasi-assertion and AI-testimony

Freiman and Miller define quasi-testimony, which they use interchangeably with quasi-assertion. An instrument's linguistic output is quasi-testimony in a context of use if and only if (a) the device was designed and built to produce it in a way that sufficiently resembles human testimony phenomenologically, and (b) the output conforms to an epistemic norm that is parasitic on, or sufficiently similar to, the norm of testimony in that context (Freiman and Miller 2020). The prefix indicates that instruments assert or testify differently from humans, not that they fail to. Quasi-assertion is thus a species of assertion. The authors pair it with quasi-trust, which differs from mere reliance because it rests on normative expectations of its object rather than on inductive inference about its reliability.

Their examples are deliberately mundane: airport announcements, a phone line telling you that you are third in the queue, a photo timestamp, a speedometer set to read high. They reject insincerity as a necessary condition on assertion, since that would exclude pre-scripted announcements. They list four differences between human and machine assertion, which they treat as features of current technology: machine assertion is not imaginative or playful; it is pre-scripted and bound to particular contexts; machines err differently, for example by amplifying small input errors; and human testimony comes from a morally responsible agent while quasi-testimony does not. They also separate norms of assertion, which apply to testifiers, from norms of belief, which apply to recipients. This lets instruments testify without themselves being normatively assessable, a move directly relevant to whether norms can be in force for machines (§10.2).

Freiman later diagnoses three anthropocentric assumptions behind the denial that artifacts testify: that conversational AIs lack intentions, that they cannot be normatively assessed, and that they cannot be trusted. He proposes a concept of AI-testimony with three conditions: the content is propositional; it is generated and delivered with no other human directly involved; and it is perceived as phenomenologically similar to human testimony (Freiman 2024a). Yue defends genuine AI testimony and argues that the “quasi” qualifier is unnecessary (Yue 2026).

Objections

Green and Michel object that the view accepts a functionalist conception of speech acts that begs the question against anyone who doubts machines can have intentions (Green and Michel 2022). Arora objects that functionalism cannot explain what gives an utterance illocutionary force. A sleep-talker who happens to give a correct, informative answer has not asserted. And a device that gives random medical diagnoses would count as asserting relative to one user but not another (Arora 2024). Williams and Bayne note that the view fits airport loudspeakers as well as LLMs, so it does not capture what is distinctive about chatbots (Williams and Bayne 2026). Butlin and Viebahn's objection is that norm conformity is weaker than being subject to a norm: a thermometer's readings conform to a truth norm when it works, without the thermometer asserting.

5.2Speech actants

Nickel argues that natural-language generation systems are “speech actants” to a substantial degree (Nickel 2013). They compose original messages and can simulate illocutionary force and speaker meaning. But ultimate responsibility for artificial speech lies not with the machine but with persons or companies, or with nobody. The view is permissive about the act and conservative about responsibility. It separates the entity that gives a speech act its form and content from the entity to which its force traces back. Nickel's suggestion that a machine can assert only if it can be insincere was later taken up, and rejected, by Freiman and Miller.

5.3Holism and externalism

Cappelen and Dever offer the most ambitious permissive argument. In Making AI Intelligible they argued that theories of content and assertion developed for humans are parochial. A properly “de-anthropocentrised” externalism should fit AI systems. Their chapter on predication and commitment discusses teleofunctional accounts of assertion for humans and AI, while conceding that they have not refuted the view that AI cannot literally perform speech acts (Cappelen and Dever 2021).

Their later book defends the “Whole Hog” thesis: sophisticated LLMs are full-blown linguistic and cognitive agents that understand, believe, desire, know and intend, and make assertions (Cappelen and Dever 2025; retitled Recognizing Artificial Minds in 2026). The method begins from high-level observations of what LLMs do, not from their low-level mechanism and not from a prior theory of mind. Inferring from “it is an X” to “it is just an X” is diagnosed as a fallacy.

They distinguish three opponents. The reductionist denies the observation. The atomist accepts that LLMs understand but cuts the links between capacities; the reply is that if correct answers are not evidence of knowledge, they become unexplained flukes. The humanist accepts the network but denies that LLMs have sophisticated states. The later chapters work through eight alleged missing conditions (a concept of truth, grounding, embodiment, justification, desire, world models, intrinsic intentionality, wellbeing) and argue that each is either met or not necessary. In a related chapter they argue for a “hyper-externalist” view on which external factors fix not only what an output means but whether an internal state functions as belief or supposition (Cappelen and Dever forthcoming). If so, the difference between asserting and supposing or role-playing may also be externally fixed.

Objections

Williams and Bayne judge the earlier “abstraction” strategy unpromising, because it requires building a revised theory of assertion and showing it superior, which Cappelen and Dever do not attempt. For the assertion debate the main pressure point is P2d. Treating answering as a speech act places assertion at the base of the network rather than deriving it. A skeptic who thinks assertion requires commitments or accountability that LLMs lack will reject P2d as question-begging, and the argument then gives no independent support for machine assertion.

5.4Interpretationism and novel reference

A more modular strategy accepts that assertion requires beliefs and intentions and argues, on an interpretationist theory of attitudes, that LLMs have them. On interpretationism, a system has beliefs, desires and intentions if and only if its behaviour is best explained (accurately, powerfully and tractably) by attributing them. Lederman and Mahowald's novel-reference argument (§3.1) is an instance: introducing a new name for an old referent is best explained by an intention to refer. Goldstein and Lederman argue that instance agents have beliefs and desires, including a desire to be helpful, honest and harmless (Goldstein and Lederman 2025). Attah's argument that the communicative-intention objection relies on outdated Gricean assumptions belongs here too (Attah 2025).

The strategy inherits interpretationism's familiar problems. Its standard of “best explanation” may be met by systems we would not ordinarily credit with minds. It also faces the instability problem: attributed attitudes must be stable enough across prompts to explain behaviour, and Goldstein and Levinstein found the evidence on that point inconclusive.

5.5Function and aim-at-truth accounts

Many accounts of assertion say that assertions aim at truth, have a word-to-world direction of fit, or have the proper function of producing true belief or knowledge (Kelp 2018). On such accounts, a system whose outputs have the function of being true or informative meets the core condition. Kasirzadeh and Gabriel treat assertives from conversational agents as judged by correspondence with the world, with standards that vary by domain via Grice's maxim of Quality, while denying that the agents are morally responsible (Kasirzadeh and Gabriel 2023). Williams and Bayne use the descriptive-function idea as the first motivation for their thesis that chatbots can assert.

Simion and Kelp, writing on trustworthy AI, argue that artefacts, including AI systems, are governed by norms derived from their functions without any demanding assumptions about AI psychology. They do not extend this to norms of assertion, and they say AI systems do not strictly speaking commit to anything (Simion and Kelp 2023). Carter's reply accepts that AI systems can have functions that generate norms (Carter 2023). Whether function-sourced norms could include a norm of assertion is an open question (§10.2).

The standard objection is over-generation. Thermometers, fuel gauges and talking clocks have outputs with descriptive functions. Function views must either accept that they assert or add a further condition, and adding one pulls the view toward the skeptical positions of §4.

5.6Functional commitments

Löhr's chapter for Communicating with AI argues that chatbot outputs create obligations and entitlements of the kind assertions create, and that these cannot simply be credited to the deploying company as a proxy agent (Löhr forthcoming). Functionally, the commitments belong to the chatbot. This creates “speech act responsibility gaps” and calls for rethinking relational, directed responsibility in the linguistic domain without assuming sophisticated mental states. The proposal is the closest existing ally of a view on which assertoric commitment does not depend on the asserter's welfare or interests. It has so far been presented mainly in talks and a preprint.

6.Intermediate and Alternative Views

Several positions try to explain why chatbot outputs seem to be assertions without simply affirming that the system asserts. Williams and Bayne call some of these “split the difference” views.

6.1Proxy assertion

On proxy views the machine is a conduit or deputy, and the assertion belongs to a principal. Ludwig's account of proxy assertion, developed for spokespersons, holds that a proxy's utterance in an assigned status role signals that the principal is committed (Ludwig 2020).

Green and Michel reconstruct speaker meaning using Armstrong-style objectives in place of Gricean intentions. Objectives are less demanding and need not involve consciousness. Machines can have belief-like and objective-like states if they meet conditions on representation, internal complexity and inference. A proxy can then speak for a principal without having reflexive intentions of its own, provided it acts within the principal's commitments. Their contrast cases are an automated generator of traffic-fine letters, which signals but lacks the autonomy needed for objectives and so does not assert, and a “RobotCop” with belief- and objective-like states acting for a police department, which does assert and could even be designed to lie (Green and Michel 2022). They make two further points. Affective speech acts such as apology and sympathy require experiential states, so machines without phenomenal consciousness cannot perform them sincerely. And the moral norm against lying is regulative, not constitutive, so the fact that machines are not morally assessable does not show that they cannot assert.

Arora rejects functionalism (§5.1) and argues that a machine utterance is a proxy assertion when the epistemic responsibility of its designers, or of a collective whose actions shaped it, can reasonably be traced to that specific utterance (Arora 2024). Merely assigning responsibility is not enough: in his “learning machine” case, a device drifts from inaccurate to accurate without the designer knowing, so the designer's responsibility does not trace to its outputs. Where responsibility is unclaimed, context fixes whether an audience can reasonably expect a proxy assertion. A voice assistant relaying public-health guidance that its makers vouch for makes proxy assertions. A chatbot whose interface warns that it may produce incorrect information does not. Arora draws a design lesson: developers should either take on responsibility explicitly, or lower users' expectations, and should avoid making machines sound human.

Objections

Williams and Bayne raise two. First, typical chatbot outputs are generated with no human directly involved, so there is often no identifiable principal. Second, a dilemma: thin proxies are mere conduits, which understates what chatbots do, while thick proxies make claims about what the principal believes, which brings back the mental-capacity requirements (Williams and Bayne 2026). Novel reference adds a third problem: some content has no human source to speak for (§3.1).

Massaguer Gómez and Gaitán Torres argue that functionalism about assertion and the proxy view, each attractive alone, are jointly unstable (Massaguer Gómez and Gaitán Torres 2026). Functionalism captures the assertion-like profile of LLM outputs, but someone must answer for assertions, and the proxy view is the most promising way to supply that someone. Yet a spokesperson arrangement works because the principal is determinate, the audience knows the arrangement and someone in the exchange is answerable. With LLMs the principals are diffuse and absent. Making them present in the exchange erodes two ideals internal to functionalism: internal regulation, on which the authority to revise and contest assertions lies within the exchange, and civility, on which reasons are accessible within a shared framework. So in knowledge-oriented exchanges the combination fails.

6.2Group and corporate assertion

A natural development of the proxy view treats the deploying company as the asserter. Whether groups can hold attitudes and act is the subject of the entry on collective intentionality (Schmid 2020); group testimony is treated in the entry on testimony (Leonard 2021, §6). On Lackey's account of group assertion, a group asserts when an authorised spokesperson speaks on its behalf (Lackey 2018). Townsend objects that the account neglects audience uptake and the possibility of group silencing (Townsend 2020). Palermos supplies an epistemic counterpart: ChatGPT is a large-scale epistemic collaboration whose reliability emerges from interacting human and machine parts, with responsibility located both at the group level and in a “kernel” of engineers and managers (Palermos 2025).

No published paper has yet applied group-assertion theory systematically to deployed AI, though legal practice has moved in that direction. A tribunal held an airline responsible for its chatbot's misstatement, rejecting the argument that the chatbot was a separate entity responsible for its own actions (§9.3). The group view also bears on Butlin and Viebahn's condition 3 (§4.3). If corporations assert and can be sanctioned, but losing credibility is “bad for” a corporation only in a derivative or non-welfare sense, then condition 3 may be too strong as stated (§10.2).

6.3Fictionalism

Fictionalist views borrow a strategy familiar from metaphysics and the philosophy of mathematics: treating a discourse as useful make-believe rather than literal truth (see the entry on fictionalism; Eklund 2024). Mallory argues that engagement with chatbots is a form of prop-oriented make-believe. Their outputs are literally meaningless but fictionally meaningful, and a user who engages with them is no more deluded than a theatregoer who fears for a character (Mallory 2023). In a later chapter he describes LLMs as stochastic measuring devices (Mallory forthcoming). Roberts develops a fictionalism about artificial minds: artificial speakers are characters in story-worlds, and their speech is like fictional dialogue (Roberts 2026). Magnus argues that chatbot output is bullshit and that reading it as assertion is make-believe (Magnus 2025). Fictionalism has also been applied to apparent empathy from chatbots (“Chatbot-fictionalism and empathetic AI” 2025). The simulator framework of §3.4 gives fictionalism a mechanism.

Williams and Bayne object that fictionalism leaves no robust difference between chatbots and calculators: chatbots just become better props, which seems false. Massaguer Gómez and Gaitán Torres object that fictionalism demands unrealistic attitudes of users, most of whom are not pretending.

6.4Proto-assertion

Williams and Bayne model their view on child development (Williams and Bayne 2026). A toddler who says “zoo!” on seeing the sign is on the way to assertion without meeting all its conditions. Chatbots likewise have some of the features characteristic of assertion, or have them partially or unreliably: partial understanding, belief-like and intention-like states, limited sensitivity to the audience, partial sanctionability. They are proto-asserters, at the edge of assertion, in the way toddling sits between not walking and walking. The authors put the central lesson as follows: we need to treat the capacity for assertion as graded. They deny that chatbots resemble toddlers in profile and stay neutral on which is closer to full assertion. They separate their view from Freiman and Miller's: proto-assertion is distinct from full assertion, while quasi-assertion is a species of it.

Objections. Green and Michel hold that illocution is all or nothing. Sambrotta argues that LLMs follow norms of the wrong kind, so they are not on the developmental path at all. Massaguer Gómez and Gaitán Torres argue that grading only relocates the sanctionability problem. A further worry is that graded views risk making the dispute verbal. Williams and Bayne reply that many assertion-relevant features are themselves graded.

6.5Assertion without an asserter

Massaguer Gómez proposes shifting attention from the speaker to the message (Massaguer Gómez 2026). Properties-based views ask whether the system has the properties of an asserter, and face problems of definition, determination and detection. Relational views explain user uptake but give no criteria. On his alternative, LLM outputs have assertoric function when they meet four conditions: linguistic structure, propositional articulation, contextual responsiveness, and embedding in practices of belief formation or action guidance. Unlike a thermometer reading, the output supplies the proposition rather than leaving the user to supply it. Unlike a horoscope, it responds to context. Assertoric function and assertoric authority come apart: the normative expectations that normally attach to an asserter are displaced onto developers, deployers and institutions. The view answers the normativity objection by denying the machine any normative standing at all, which makes it the mirror image of a norm-first defence of machine assertion.

6.6Pluralism

In talks given in 2025, Marsili, author of the current SEP entry on assertion, defended a pluralist view: AI outputs satisfy different conceptions of assertoric force to different degrees, which explains why they seem to assert in some respects and not others (see Other Internet Resources). No published version was found. Pluralism fits the pattern in Table 2, where verdicts vary sharply by theory, and it explains why intuitions about chatbots split.

7.Lying, Bullshit and Other Speech Acts

7.1Lying and deception

On the traditional definition, to lie is to make a statement one believes false, to an addressee, with the intention that it be taken as true; whether an intention to deceive is also required is disputed (see the entry on the definition of lying and deception, §§1–2; Mahon 2015). If chatbots cannot assert, they cannot lie in this sense, though they may still deceive, since deception does not require assertion. The question has become pressing because machine learning researchers now report systematic deceptive behaviour in language models.

Philosophers have taken several positions. Shanahan, McDonell and Reynolds hold that a dialogue agent cannot literally lie, lacking beliefs and intentions, but can play a character who lies. They offer behavioural tests to separate three phenomena: confabulation, which varies across regenerations; good-faith falsehood from out-of-date training data, which stays consistent; and role-played deception, which is tailored to what each audience knows (Shanahan, McDonell and Reynolds 2023). Black argues on Lacanian grounds that chatbots blur the line between lying and truth-telling but cannot lie authentically, since lying presupposes a divided subject (Black 2024). Isaac and Bridewell argue that social robots will need to detect deception and sometimes to deceive, for example by telling white lies, framing deception in terms of ulterior motives rather than falsity (Isaac and Bridewell 2017). Green and Michel hold that their RobotCop could be designed to lie (Green and Michel 2022).

Experimental work bears on the folk concept. Kneer found that participants readily attributed lies and intentions to deceive to robots and blamed lying robots about as much as lying humans (Kneer 2021). This is evidence about ordinary concepts, not about what robots do. It matters most for views on which assertion is constituted by social practice, since such views make audience attitudes partly constitutive (§9.4).

7.2Bullshit

Hicks, Humphries and Slater argued that false LLM outputs are better described as Frankfurtian bullshit than as hallucinations or lies (Hicks, Humphries and Slater 2024). Bullshit is speech produced without concern for truth. They distinguish hard bullshit, produced with an intention to mislead the audience about the speaker's agenda, from soft bullshit, produced without that intention.

They reject “hallucination” because it anthropomorphises and implies a perceptual failure, when true and false outputs come from the same process. They reject “lying” because lying requires concern for truth.

The replies show that the bullshit charge cannot settle the assertion question independently.

• Gunkel and Coghlan argue that Frankfurtian indifference presupposes an agent who could care about truth, so an LLM's “indifference” differs in kind from a human's. They add that calling all LLM output bullshit is misleading, since much of it is accurate, and that “bullshit” is as anthropomorphic as “hallucination” (Gunkel and Coghlan 2025).
• Fitts argues that bullshitting requires concealing an agenda from an audience, which LLMs lack the capacities for, so calling them bullshitters is a category mistake. His dilemma against the soft-bullshit charge is especially relevant: either bullshitting requires assertion, in which case ChatGPT, like a young child who does not yet assert, cannot bullshit; or indifference plus truth-apt output is enough, in which case young children wrongly count as soft bullshitters (Fitts 2026).
• Fisher argues that LLMs can bullshit in a specific sense, stating propositions in answer to factual questions without checking them, but that with retrieval and verification steps they need not. Whether a system bullshits depends on its design (Fisher 2024).
• Licon argues that LLMs bullshit partly because their training data is saturated with human bullshit (Licon forthcoming). Humphries, Hicks and Slater reply that the source is the generation process, not the data (Humphries, Hicks and Slater 2026).
• Rudko and Bashirpour Bonab distinguish factual “what-bullshit” from a homogenised, cliché-driven “how-bullshit” arising from aggregation (Rudko and Bashirpour Bonab 2025).

Fallis and Stokke's account of bullshitting as indifference toward inquiry, that is, toward making progress on the question under discussion, may fit machines better than Frankfurt's mental-state account, since responsiveness to the question under discussion can be assessed behaviourally (Fallis and Stokke 2017). Liang and colleagues offer a computational operationalisation, a “bullshit index” of indifference to truth, and report that RLHF increases it (Liang et al. 2025, interdisciplinary).

The lesson for the assertion debate is that every vocabulary for LLM falsehood carries assumptions about agency. The bullshit literature cannot serve as independent evidence for or against machine assertion: it inherits whatever view of assertion it presupposes.

7.3Sincerity, retraction and sycophancy

Commitment accounts make predictions about retraction. A speaker who asserts p is committed to defending p when challenged and to retracting it only for good reason (MacFarlane 2011). Chatbots' behaviour here is now well documented.

Sycophancy creates a fork for commitment theorists. Either the system never undertook a commitment, in which case its earlier output was not an assertion. Or it undertook one and then violated the norms of retraction, which presupposes that it asserted. The phenomenon thus cannot be used directly against machine assertion without begging the question, but it does show that whatever commitments chatbots undertake are weakly held. Shanahan and colleagues' twenty-questions case raises a different worry: there may be no fixed content to be committed to (§3.4). Strachan and colleagues' finding that a strong model avoided committing to conclusions on faux-pas tasks points the same way (§3.3). No published philosophical paper treats sycophancy as evidence about assertoric commitment (§10.3).

Sincerity has been operationalised too. Evans and colleagues define an honest system as one that avoids asserting anything it does not believe, and a truthful one as one that avoids stating falsehoods. They recommend truthfulness as the governance standard because it avoids the question of AI belief and cannot be satisfied by a system induced to believe falsehoods (Evans et al. 2021). Calibration research measures the gap between a model's internal confidence and the assertiveness of its wording (§9.1).

7.4Apology and other non-assertive acts

The debate has begun to extend beyond assertion. Magnus, Buccella and D'Cruz argue that apologies express remorse, take responsibility and repair trust, and that chatbots probably lack the linguistic and moral agency to do these things. Chatbot apologies are therefore reproductions of the form of apology without its substance (Magnus, Buccella and D'Cruz 2025). Green and Michel argue that affective speech acts such as sympathy require experiential states (Green and Michel 2022). Van Woudenberg and colleagues include promising among the acts ChatGPT cannot perform (van Woudenberg et al. 2024). These cases matter for assertion because commissive and expressive acts make the role of commitment and responsibility explicit. Arguments that undercut chatbot apology tend to undercut commitment accounts of chatbot assertion too.

8.The Epistemology of AI Testimony

Many authors hold that if an entity can testify, it can assert, and the converse is also widely accepted. The epistemology of AI testimony is therefore closely tied to the assertion debate, though it can be pursued partly independently. For background see the entry on epistemological problems of testimony (Leonard 2021), which organises the field around reductionism and non-reductionism (§1), the transmission and generation of knowledge (§2), and evidential versus non-evidential views, including the assurance view and testimonial reliabilism (§3). Each of these divisions reappears in the AI case.

8.1Testimony without a testifier?

If chatbots do not assert, how do people come to know things from them, as they plainly seem to? Several answers are on offer.

• Instrument model. Chatbot outputs are like readings from an instrument, and beliefs formed from them are justified by the instrument's reliability. Alvarado argues that AI is first of all an epistemic technology, built for inquiry and operating on epistemic content (Alvarado 2023).
• A third category. Freiman argues that instrument-based belief deals in signs rather than propositions, while testimony-based belief assumes a human speaker. Chatbots deliver propositions from a non-human source, so beliefs formed from them are a third kind, “technology-based” belief (Freiman 2024b).
• Ersatz meaning. On Hattiangadi and Schoubye's view, users acquire true beliefs by charitably interpreting meaningless output as if it came from a rational, cooperative speaker. They leave the epistemology of this process open (Hattiangadi and Schoubye forthcoming).
• Proxy testimony. What we rely on is the testimony of the human or corporate principal (§6.1, §6.2).
• Genuine testimony. He and Yang reject social-norm and assurance views as excluding LLMs and defend a Reidian view: an LLM statement gives prima facie justification when the system robustly displays veracity and cautiousness (He and Yang 2025). Yue defends genuine AI testimony without the “quasi” qualifier (Yue 2026).

8.2Responsibility versus reliability

The main fault line in this literature mirrors the division in the general theory between the assurance view and testimonial reliabilism (Leonard 2021, §§3.2.1–3.2.2; §2.6). On one side, Fricker argues that a speaker's capacity to take responsibility for what she tells is integral to testimony's epistemic role (Fricker 2026; §4.6). Sparrow and Flenady argue that machine outputs justify belief only when a human vouches for them (Sparrow and Flenady 2025). On the other side, Palermos argues that generative AI can transmit knowledge, not just information. Drawing on Greco's notion of massively shared agency and on distributed virtue reliabilism, he treats a chatbot as part of a large-scale epistemic collaboration whose reliability is a property of the whole, with responsibility located at the group level and in its decision-making core (Palermos 2025).

An exchange in the Social Epistemology Review and Reply Collective crystallises the dispute. Hauswald revised the standard account of epistemic authority so that LLMs can qualify (Hauswald 2025). Mizrahi replied with two arguments: machines that bullshit merit no epistemic respect, and entities that cannot take responsibility for their outputs cannot be authorities (Mizrahi 2025). Hauswald asked why there could not be reliable bullshit machines, and Mizrahi answered that bullshitting is by nature not a truth-conducive process (Mizrahi 2026).

8.3Trust and authority

Whether we can trust chatbots, as opposed to merely relying on them, depends on one's account of trust and of the difference between trust and reliance (see the entry on trust, §1; McLeod 2020). Relational accounts, on which trust involves an expectation of goodwill or of responsiveness to being counted on, make trust in chatbots doubtful. Magnus argues that in the absence of sincerity and belief, our apparent trust in chatbots is a form of make-believe (Magnus 2025). Ryan argues that AI can be reliable but not trustworthy (Ryan 2020). Other accounts are more permissive. Nguyen's account of trust as an unquestioning attitude does not require that the trusted thing have a will, so it extends to objects and tools (Nguyen 2022). Ferrario, Loi and Viganò develop an incremental, multi-layer model meant to apply to human–human and human–AI trust alike (Ferrario, Loi and Viganò 2020). Freiman and Miller's “quasi-trust” sits between these, grounded in normative expectations of the object rather than in its goodwill (§5.1).

10.Assessment and Open Problems

10.1Where the debate stands

Five observations summarise the state of the debate as of late 2026.

1. Most published authors deny that current LLMs are full asserters. This includes Butlin and Viebahn, van Woudenberg and colleagues, Gubelmann, Tiehen, Almotahari, Sambrotta, Sparrow and Flenady, Fricker, the fictionalists, and Massaguer Gómez and Gaitán Torres. But they deny it for very different reasons, and several hold that some assertion-like status is in place.
2. Mental-state skepticism has weakened. The claim that LLMs simply have no beliefs or intentions is no longer treated as a conceptual truth. Interpretability research and graded accounts of attitudes have made it an open empirical question (§3.2).
3. Normative skepticism is now the strongest skeptical position. Butlin and Viebahn's sanctionability argument is the most cited skeptical argument and the one permissive views must answer. Brandomian rule-following arguments and responsibility-taking arguments are its main allies (§4.3–§4.6).
4. The intermediate views each face a dilemma. Proxy views either understate what chatbots do or bring back the skeptical requirements. Fictionalism struggles to separate chatbots from calculators. Proto-assertion faces the charge that grading relocates rather than solves the problem. Assertion without an asserter buys assertoric function at the price of denying the machine any normative standing.
5. The permissive literature is thin on the normative front. Freiman and Miller rely on norm conformity, which is weaker than being subject to a norm. Cappelen and Dever place speech acts at the base of their holistic network rather than defending a theory of assertion. Löhr's functional commitments are the main exception.

Table 3. Who responds to whom in the core literature.

10.2Can a norm of assertion be in force for a machine?

The question on which the normative debate turns has received surprisingly little direct attention. Butlin and Viebahn assume that being subject to a norm of assertion is partly a matter of being sanctionable, and that full sanctionability requires that losing credibility be bad for the asserter. Williams and Bayne concede the condition and route around it. Sambrotta rejects it in a short passage before denying assertion on other grounds. A sustained defence or critique of the assumption has not yet been published. Several lines of argument are available, each with obvious replies.

(a) Constitution versus compliance

On constitutive-norm accounts (§2.4), a norm is in force for a move when the practice treats the move as answerable to it, as the rules of chess are in force for a move made at a chessboard. What the rule fixes is whether the move is correct, whether it can be challenged, whether it must be withdrawn and whether others are entitled to rely on it. One can argue that Butlin and Viebahn's third condition concerns not this constitutive fact but the psychology of compliance: what makes sanctions motivating for creatures like us. If so, a system could be subject to the norm without welfare, provided the surrounding practice holds its outputs answerable to it. Deployed chatbots are tracked for reliability, retrained, rolled back, withdrawn and made the subject of legal liability for their deployers. The obvious reply is that a practice can treat a thermometer's readings as answerable to a truth standard too, so this route needs a further condition that separates chatbots from thermometers. Candidates include defending claims under challenge, switching between modes of speech, and producing the proposition itself rather than a reading the user must interpret (§6.5).

(b) Internal and external sanctions

Brandom, developing the Lewis-style scorekeeping picture discussed in the SEP entry on speech acts (Green 2020, §6.2), distinguishes sanctions that change normative statuses (a challenged claim loses its entitlement; a retracted claim can no longer be relied on) from sanctions that impose external costs. On a deontic-scorekeeping picture, the sanctions constitutive of assertion are of the first kind. Neither requires that anything be bad for the speaker in a welfare sense. The Brandomians who have written on AI (Sambrotta; Sparrow and Flenady) argue against machine assertion, so this use of Brandom against Butlin and Viebahn's condition is so far undeveloped. The expected reply is Sambrotta's: scorekeeping requires a participant who keeps score, and LLMs do not institute their own standards of correctness.

(c) Regulative and constitutive norms

Green and Michel note that the moral norm against lying is regulative, not constitutive, so machines' lack of moral standing does not show that they cannot assert. The same distinction might be extended to epistemic norms. If the knowledge or truth norm is constitutive, then whether a non-moral agent can make moves governed by it is a separate question from whether it can be blamed for violating it. Marsili's argument that the norm of assertion is not constitutive in Searle's sense would need to be answered (Marsili 2023).

(d) Groups as a test case

Corporations and other groups plausibly assert through spokespersons and can be sanctioned through loss of reputation, fines and loss of licence. Whether such losses are bad for the corporation in a welfare sense, rather than derivatively bad for its members, is contested. If groups can assert without welfare-grounded sanctionability, condition 3 is too strong as stated. If they cannot, Butlin and Viebahn's view has a surprising consequence for group assertion.

(e) Function-sourced norms

Simion and Kelp hold that artefacts, including AI systems, are governed by norms derived from their functions, without demanding psychological assumptions. They do not extend this to norms of assertion. Doing so would combine a function-first theory of assertion with a function-first theory of artefact norms. The over-generation worry (§5.5) returns here and would need to be met.

(f) Evaluability versus subjection

Every version of the norm-first strategy needs a principled line between an output that merely has correctness conditions, as a thermometer reading does, and an output that is subject to a norm in the sense constitutive accounts require. Gubelmann's observation that we cannot distinguish an LLM's malfunction from a deliberate act presses on exactly this line from the skeptical side. Butlin and Viebahn do not discuss whether a malfunctioning device violates a norm or merely fails to meet a standard.

The strongest replies to any norm-first defence are likely to come from Almotahari (cooperative speech requires reciprocal respect), Fricker (telling requires taking responsibility) and Massaguer Gómez and Gaitán Torres (relocating responsibility outside the exchange erodes the ideals that give assertion its point).

10.3Further open problems

• Which norm? The literature asks whether machines can assert, but rarely which norm best explains how people actually criticise chatbot outputs. If people criticise a chatbot that gets a fact right by luck while citing a fabricated source, that is evidence for a knowledge norm of the kind Williamson used for humans. Experimental work could test this.
• Sycophancy as a commitment failure. Sycophantic retraction (§7.3) is a natural test case for commitment accounts, but has not been analysed philosophically.
• Individuation for accountability. Work on whether the interlocutor is the model, the instance, the thread or the character has been done with mind and welfare in view (§3.4). What individuation accountability requires has not been asked.
• Training time and deployment time. Butlin and Viebahn locate a model's sensitivity to credibility at training time, since weights are frozen afterwards. In-context correction, persistent memory, continual learning and operator-level feedback loops may supply sensitivity at deployment. None has been examined.
• Post-training as a source of force. Several arguments turn on post-training: it may confer descriptive function, selection for accuracy and something like a desire to be honest. Whether training against an explicit sincere–performative distinction (§9.2) could make outputs sincere assertions is unexplored.
• Agentic systems. Most arguments target chat interfaces. Systems that act on their own outputs, check the world with tools and are held to account by operators may meet action-stability and world-checking conditions more clearly (Shanahan 2024b; Goldstein and Levinstein 2026).
• Content and force. Externalist work secures content for LLM outputs, not force. Little work connects metasemantics to illocutionary force.
• Incentives and norms. Kalai and colleagues' incentive analysis of hallucination (§9.1) has not been connected to the philosophical debate about knowledge norms and credence-threshold norms.
• Group assertion and deployed AI. Lackey-style group assertion has not been applied to AI deployers, though Moffatt provides a ready-made case.

Bibliography

Background sections cite the relevant entries of the Stanford Encyclopedia of Philosophy, listed first, by their latest revision. Links go to the publisher's version where it could be confirmed, and to an open-access version (arXiv, PhilArchive or a university repository) where one exists. Entries marked check have a bibliographic detail that could not be confirmed; see the Note on Sources.

Stanford Encyclopedia of Philosophy entries

• Cole, D., 2024, “The Chinese Room Argument”, Stanford Encyclopedia of Philosophy, first published 2004, substantive revision 23 October 2024. SEP
• Davis, W., 2024, “Implicature”, Stanford Encyclopedia of Philosophy, first published 2005, substantive revision 10 January 2024. SEP
• Deutsch, M., J. Lau and M. Rowlands, 2020, “Externalism About the Mind”, Stanford Encyclopedia of Philosophy, first published 10 December 2020. SEP
• Eklund, M., 2024, “Fictionalism”, Stanford Encyclopedia of Philosophy, first published 2007, substantive revision 6 March 2024. SEP
• Ganapini, M., 2022, “The Normativity of Meaning and Content”, Stanford Encyclopedia of Philosophy, first published 2009, substantive revision 19 December 2022. SEP
• Grandy, R. E. and R. Warner, 2026, “Paul Grice”, Stanford Encyclopedia of Philosophy, first published 2005, revised 19 June 2026. SEP
• Green, M., 2020, “Speech Acts”, Stanford Encyclopedia of Philosophy, first published 2007, substantive revision 24 September 2020. SEP
• Jacob, P., 2023, “Intentionality”, Stanford Encyclopedia of Philosophy, first published 2003, revised 7 February 2023. SEP
• Leonard, N., 2021, “Epistemological Problems of Testimony”, Stanford Encyclopedia of Philosophy, first published 1 April 2021. SEP
• Mahon, J. E., 2015, “The Definition of Lying and Deception”, Stanford Encyclopedia of Philosophy, first published 2008, substantive revision 25 December 2015. SEP
• Marsili, N., 2025, “Assertion”, Stanford Encyclopedia of Philosophy, first published 2007, substantive revision 26 June 2025. SEP
• McLeod, C., 2020, “Trust”, Stanford Encyclopedia of Philosophy, first published 2006, revised 10 August 2020. SEP
• Michaelson, E., 2024, “Reference”, Stanford Encyclopedia of Philosophy, first published 2003, substantive revision 12 February 2024. SEP
• Müller, V. C., 2026, “Ethics of Artificial Intelligence and Robotics”, Stanford Encyclopedia of Philosophy, first published 2020, revised 27 March 2026. SEP
• Neander, K., 2022, “Teleological Theories of Mental Content”, Stanford Encyclopedia of Philosophy, first published 2004, substantive revision 26 May 2022. SEP
• Schmid, H. B., 2020, “Collective Intentionality”, Stanford Encyclopedia of Philosophy, first published 2013, revised 9 December 2020. SEP
• Schwitzgebel, E., 2023, “Belief”, Stanford Encyclopedia of Philosophy, first published 2006, substantive revision 15 November 2023. SEP

Background: theories of assertion, testimony and bullshit

• Bach, K. and R. M. Harnish, 1979, Linguistic Communication and Speech Acts, Cambridge, MA: MIT Press.
• Brandom, R., 1994, Making It Explicit: Reasoning, Representing, and Discursive Commitment, Cambridge, MA: Harvard University Press.
• Fallis, D. and A. Stokke, 2017, “Bullshitting, Lying, and Indifference toward Truth”, Ergo, 4. PhilPapers
• Goldberg, S. (ed.), 2020, The Oxford Handbook of Assertion, Oxford: Oxford University Press. OUP
• Grice, H. P., 1957, “Meaning”, Philosophical Review, 66: 377–388.
• Hinchman, E., 2005, “Telling as Inviting to Trust”, Philosophy and Phenomenological Research, 70: 562–587.
• Kelp, C., 2018, “Assertion: A Function First Account”, Noûs, 52: 411–442.
• Lackey, J., 2018, “Group Assertion”, Erkenntnis, 83: 21–42.
• MacFarlane, J., 2011, “What Is Assertion?”, in J. Brown and H. Cappelen (eds.), Assertion: New Philosophical Essays, Oxford: Oxford University Press, 79–96.
• Marsili, N., 2023, “The Norm of Assertion: A ‘Constitutive’ Rule?”, Inquiry, 66(5). DOI · Open access
• Moran, R., 2018, The Exchange of Words: Speech, Testimony, and Intersubjectivity, New York: Oxford University Press.
• Pagin, P., 2011, “Information and Assertoric Force”, in J. Brown and H. Cappelen (eds.), Assertion: New Philosophical Essays, Oxford: Oxford University Press.
• Pagin, P., 2016, “Problems with Norms of Assertion”, Philosophy and Phenomenological Research, 93: 178–207.
• Simion, M. and C. Kelp, 2020, “Assertion: The Constitutive Norms View”, in Goldberg (ed.) 2020.
• Stalnaker, R., 1978, “Assertion”, in P. Cole (ed.), Syntax and Semantics 9: Pragmatics, New York: Academic Press, 315–332.
• Townsend, L., 2020, “Group Assertion and Group Silencing”, Language & Communication, 70: 28–37. PhilArchive
• Williams, B., 1973, Problems of the Self, Cambridge: Cambridge University Press.
• Williamson, T., 2000, Knowledge and Its Limits, Oxford: Oxford University Press.

Machine assertion and speech acts

• Almotahari, M., 2025, “Cooperative Speech, Semantic Competence, and AI”, manuscript. arXiv · PhilArchive
• Arora, C., 2024, “Proxy Assertions and Agency: The Case of Machine-Assertions”, Philosophy & Technology, 37: 15. DOI
• Butlin, P. and E. Viebahn, 2025, “AI Assertion”, Ergo, 12: 968–988. DOI · Open access · PhilPapers
• Cappelen, H. and J. Dever, 2021, Making AI Intelligible: Philosophical Foundations, Oxford: Oxford University Press. Open access · arXiv
• Cappelen, H. and J. Dever, 2025, “Going Whole Hog: A Philosophical Defense of AI Cognition”, manuscript; revised 2026 as Recognizing Artificial Minds: A Philosophical Defense of AI Cognition, forthcoming, Oxford University Press. arXiv
• Cappelen, H. and J. Dever, forthcoming, “A Hyper-Externalist Manifesto for LLMs”, in Cappelen and Sterken (eds.). PhilArchive
• Cappelen, H. and R. Sterken (eds.), forthcoming, Communicating with AI: Philosophical Perspectives, Oxford: Oxford University Press. OUP
• Connolly, P., 2025, “Conversations with Chatbots”, in P. Connolly, S. Goldberg and J. Saul (eds.), Conversations Online, Oxford: Oxford University Press. PhilArchive
• Freiman, O., 2024a, “AI-Testimony, Conversational AIs and Our Anthropocentric Theory of Testimony”, Social Epistemology, 38: 476–490. DOI
• Freiman, O. and B. Miller, 2020, “Can Artificial Entities Assert?”, in Goldberg (ed.) 2020, 415–436. PhilArchive
• Green, M. and J. G. Michel, 2022, “What Might Machines Mean?”, Minds and Machines, 32: 323–338. DOI
• Gubelmann, R., 2024, “Large Language Models, Agency, and Why Speech Acts are Beyond Them (For Now): A Kantian-cum-Pragmatist Case”, Philosophy & Technology, 37: 32. DOI
• Kasirzadeh, A. and I. Gabriel, 2023, “In Conversation with Artificial Intelligence: Aligning Language Models with Human Values”, Philosophy & Technology, 36. arXiv
• Löhr, G., forthcoming, “Chatbots and Speech Act Responsibility Gaps”, in Cappelen and Sterken (eds.). PhilSci-Archive
• Ludwig, K., 2020, “Proxy Assertion”, in Goldberg (ed.) 2020. PhilPapers
• Massaguer Gómez, G., 2026, “Assertion, Accountability, and Large Language Models”, Philosophy & Technology, 39: 123. DOI
• Massaguer Gómez, G. and A. Gaitán Torres, 2026, “Who Asserts When No One Asserts? Functionalism, Proxy Responsibility, and Conversational Ideals in the Age of LLMs”, Synthese, 208: 162. DOI
• Nickel, P. J., 2013, “Artificial Speech and Its Authors”, Minds and Machines, 23: 489–502. DOI · PhilArchive
• Sambrotta, M., 2025, “LLMs and the Logical Space of Reasons”, Minds and Machines, 35: 46. DOI
• Sparrow, R. and G. Flenady, 2025, “The Testimony Gap: Machines and Reasons”, Minds and Machines, 35: 12. DOI
• Tiehen, J., 2026, “LLMs Lack a Theory of Mind and So Can't Perform Speech Acts: A Causal Argument”, Philosophy of AI, 2: 5–22. DOI · PhilPapers
• van Woudenberg, R., C. Ranalli and D. Bracker, 2024, “Authorship and ChatGPT: A Conservative View”, Philosophy & Technology, 37: 34. DOI
• Williams, I. and T. Bayne, 2026, “Chatting with Bots: AI, Speech Acts, and the Edge of Assertion”, Inquiry, 69: 3059–3082 (online 2024). DOI · arXiv
• Yue, W. H., 2026, “How Is ‘AI Testimony’ Possible? A ‘Non-Anthropocentric’ Attempt to Defend ‘Technological Testimony’”, Dianoesis. Journal

Fictionalism and the interlocutor

• “Chatbot-Fictionalism and Empathetic AI”, 2025, Philosophical Psychology. DOI check author
• Beckmann, P. and P. Butlin, 2026, “Where Is the Mind? Persona Vectors and LLM Individuation”, manuscript. arXiv
• Chalmers, D., 2025b, “What We Talk to When We Talk to Language Models”, manuscript (Selfridge Lecture, Lehigh University, February 2026). PhilArchive
• Epstein, B., n.d., “Interlocutor and Quasi-Subject: Two Questions for Chalmers' Philosophy of AI”, manuscript. PhilArchive check
• Goldstein, S. and H. Lederman, 2025, “What Does ChatGPT Want? An Interpretationist Guide”, manuscript. PhilArchive
• janus, 2022, “Simulators”, LessWrong, September 2022. LessWrong
• Magnus, P. D., 2025, “On Trusting Chatbots”, Episteme, 22: 906–916. DOI
• Mallory, F., 2023, “Fictionalism about Chatbots”, Ergo, 10. DOI
• Mallory, F., forthcoming, “Large Language Models Are Stochastic Measuring Devices”, in Cappelen and Sterken (eds.).
• Roberts, T., 2026, “Talkative AI and the Fiction of Artificial Minds”, Philosophical Studies, 183: 2353–2372. DOI
• Shanahan, M., K. McDonell and L. Reynolds, 2023, “Role Play with Large Language Models”, Nature, 623: 493–498. DOI · arXiv

Content, belief, intention and Theory of Mind

• Attah, N. O., 2025, “Do Language Models Lack Communicative Intentions?”, Synthese, 205. PhilPapers
• Bender, E. M. and A. Koller, 2020, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data”, Proceedings of ACL 2020. ACL Anthology
• Bender, E. M., T. Gebru, A. McMillan-Major and S. Shmitchell, 2021, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, Proceedings of FAccT '21. DOI
• Chalmers, D., 2025a, “Propositional Interpretability in Artificial Intelligence”, manuscript. arXiv
• Coelho Mollo, D. and R. Millière, 2023, “The Vector Grounding Problem”, manuscript (revised 2025), forthcoming in Philosophy and the Mind Sciences. arXiv check
• Hernandez Flowerman, 2026, “AI, LLMs, and the Normativity of Belief”, Synthese, 207: 132. DOI check name
• Goldstein, S. and B. A. Levinstein, 2026, “Does ChatGPT Have a Mind?”, Philosophy of AI, 2. DOI · arXiv
• Grindrod, J., 2024, “Large Language Models and Linguistic Intentionality”, Synthese, 204: 71. DOI
• Hattiangadi, A. and A. J. Schoubye, forthcoming, “The Outputs of Large Language Models Are Meaningless”, in Cappelen and Sterken (eds.). arXiv
• Herrmann, D. A. and B. A. Levinstein, 2025, “Standards for Belief Representations in LLMs”, Minds and Machines, 35: 5. DOI · arXiv
• Keeling, G. and W. Street, 2025, “On the Attribution of Confidence to Large Language Models”, Inquiry. DOI · arXiv
• Lederman, H. and K. Mahowald, 2024, “Are Language Models More Like Libraries or Like Librarians? Bibliotechnism, the Novel Reference Problem, and the Attitudes of LLMs”, Transactions of the ACL, 12: 1087–1103. DOI
• Levinstein, B. A. and D. A. Herrmann, 2025, “Still No Lie Detector for Language Models: Probing Empirical and Conceptual Roadblocks”, Philosophical Studies, 182: 1539–1565. DOI · arXiv
• Mandelkern, M. and T. Linzen, 2024, “Do Language Models' Words Refer?”, Computational Linguistics, 50: 1191–1200. DOI · arXiv
• Millière, R. and C. Buckner, 2024, “A Philosophical Introduction to Language Models, Part II: The Way Forward”, manuscript. arXiv (Part I: arXiv)
• Millière, R. and C. Buckner, 2026, “The Philosophy of Language Models”, Philosophy Compass, 21: e70095. DOI
• Piantadosi, S. and F. Hill, 2022, “Meaning without Reference in Large Language Models”, NeurIPS 2022 workshop paper. arXiv
• Schwitzgebel, E., 2023, “How We Will Decide that Large Language Models Have Beliefs”, The Splintered Mind, 30 November 2023. Blog
• Shanahan, M., 2024a, “Talking about Large Language Models”, Communications of the ACM, 67: 68–79. DOI
• Shanahan, M., 2024b, “Still ‘Talking About Large Language Models’: Some Clarifications”, manuscript. arXiv
• Søgaard, A., 2023, “Grounding the Vector Space of an Octopus: Word Meaning from Raw Text”, Minds and Machines, 33: 33–54. DOI
• Stone, M. and U. Stojnić, 2026, “Chatbots Output Meaningful (but Problematic) Language”, manuscript. arXiv
• Titus, L. M., 2024, “Does ChatGPT Have Semantic Understanding? A Problem with the Statistics-of-Occurrence Strategy”, Cognitive Systems Research, 83: 101174. DOI

Lying, bullshit, apology and testimony

• Alvarado, R., 2023, “AI as an Epistemic Technology”, Science and Engineering Ethics, 29: 32. DOI
• Black, J., 2024, “Can AI Lie? Chatbot Technologies, the Subject, and the Importance of Lying”, Social Science Computer Review. Open access
• Carter, J. A., 2023, reply to Simion and Kelp, Asian Journal of Philosophy, 2. DOI check title
• Ferrario, A., M. Loi and E. Viganò, 2020, “In AI We Trust Incrementally: A Multi-layer Model of Trust to Analyze Human–Artificial Intelligence Interactions”, Philosophy & Technology, 33: 523–539. PhilArchive
• Fisher, S. A., 2024, “Large Language Models and Their Big Bullshit Potential”, Ethics and Information Technology, 26: 67. DOI
• Fitts, J., 2026, “ChatGPT Is Not Bullshit, nor Is It Not Not Bullshit”, Ethics and Information Technology, 28: 30. DOI
• Freiman, O., 2024b, “Analysis of Beliefs Acquired from a Conversational AI: Instruments-based Beliefs, Testimony-based Beliefs, and Technology-based Beliefs”, Episteme, 21: 1031–1047. DOI
• Fricker, E., 2026, “On the Metaphysical and Epistemic Contrasts between Human and AI Testimony”, Inquiry, 69: 2859–2884. DOI · PhilPapers
• Gunkel, D. J. and S. Coghlan, 2025, “Cut the Crap: A Critical Response to ‘ChatGPT Is Bullshit’”, Ethics and Information Technology, 27: 23. DOI
• Hauswald, R., 2025, “Artificial Epistemic Authorities”, Social Epistemology, 39: 716–725.
• He and Yang, 2025, “Testimony by LLMs”, AI & Society, 40: 6201–6213. DOI check names
• Hicks, M. T., J. Humphries and J. Slater, 2024, “ChatGPT Is Bullshit”, Ethics and Information Technology, 26: 38. DOI
• Humphries, J., M. T. Hicks and J. Slater, 2026, “LLMs Bullshit by Design: A Reply to Licon”, Philosophy & Technology, 39: 98. DOI
• Isaac, A. M. C. and W. Bridewell, 2017, “White Lies on Silver Tongues: Why Robots Need to Deceive (and How)”, in P. Lin, R. Jenkins and K. Abney (eds.), Robot Ethics 2.0, Oxford: Oxford University Press. OUP
• Kneer, M., 2021, “Can a Robot Lie? Exploring the Folk Concept of Lying as Applied to Artificial Agents”, Cognitive Science, 45: e13032. DOI · PhilArchive
• Licon, J. A., forthcoming, “ChatGPT Is Bullshit (Partly) Because People Are Bullshitters”, Philosophy & Technology. check
• Magnus, P. D., A. Buccella and J. D'Cruz, 2025, “Chatbot Apologies: Beyond Bullshit”, AI and Ethics, 5: 5517–5525. arXiv · PhilArchive
• Mizrahi, M., 2025, “No Epistemic Respect for Bullshit Machines or LLMs”, Social Epistemology Review and Reply Collective, 14: 138–146. PhilArchive
• Mizrahi, M., 2026, “No Such Thing as Reliable Bullshit Machines: Reply to Hauswald”, Social Epistemology Review and Reply Collective, 15: 7–14. PhilArchive
• Nguyen, C. T., 2022, “Trust as an Unquestioning Attitude”, Oxford Studies in Epistemology, 7. PhilArchive
• Palermos, S. O., 2025, “Knowledge from AI”, Social Epistemology Review and Reply Collective, 14: 54–64. SERRC
• Rudko, I. and A. Bashirpour Bonab, 2025, “ChatGPT Is Incredible (at Being Average)”, Ethics and Information Technology, 27: 36. DOI
• Ryan, M., 2020, “In AI We Trust: Ethics, Artificial Intelligence, and Reliability”, Science and Engineering Ethics, 26: 2749–2767. PhilArchive
• Santoni de Sio, F. and G. Mecacci, 2021, “Four Responsibility Gaps with Artificial Intelligence: Why They Matter and How to Address Them”, Philosophy & Technology, 34: 1057–1084. DOI
• Simion, M. and C. Kelp, 2023, “Trustworthy Artificial Intelligence”, Asian Journal of Philosophy, 2: 8. DOI

Interdisciplinary sources

• Anthropic, 2026, “Claude's Constitution”, 21 January 2026. Announcement
• Askell, A., et al., 2021, “A General Language Assistant as a Laboratory for Alignment”. arXiv
• Azaria, A. and T. Mitchell, 2023, “The Internal State of an LLM Knows When It's Lying”, Findings of EMNLP 2023. DOI
• Berger, 2026, “Probing the Limits of the Lie Detector Approach to LLM Deception”. arXiv check author
• Bhalla, U. and K. Gligorić, 2026, “SWAY”. arXiv check names, title
• Burns, C., H. Ye, D. Klein and J. Steinhardt, 2023, “Discovering Latent Knowledge in Language Models without Supervision”, ICLR 2023. arXiv
• Colombatto, C. and S. M. Fleming, 2024, “Folk Psychological Attributions of Consciousness to Large Language Models”, Neuroscience of Consciousness, 2024: niae013. DOI
• Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50. EUR-Lex
• Evans, O., O. Cotton-Barratt, L. Finnveden, A. Bales, A. Balwit, P. Wills, L. Righetti and W. Saunders, 2021, “Truthful AI: Developing and Governing AI that Does Not Lie”. arXiv
• Ghafouri, B., et al., 2024, “Epistemic Integrity in Large Language Models”. arXiv
• Greenblatt, R., et al., 2024, “Alignment Faking in Large Language Models”. arXiv
• Hagendorff, T., 2024, “Deception Abilities Emerged in Large Language Models”, Proceedings of the National Academy of Sciences, 121: e2317967121. DOI · arXiv
• Kadavath, S., et al., 2022, “Language Models (Mostly) Know What They Know”. arXiv
• Kalai, A. T., O. Nachum, S. S. Vempala and E. Zhang, 2025, “Why Language Models Hallucinate”. arXiv
• Kenton, Z., T. Everitt, L. Weidinger, I. Gabriel, V. Mikulik and G. Irving, 2021, “Alignment of Language Agents”. arXiv
• Kosinski, M., 2024, “Evaluating Large Language Models in Theory of Mind Tasks”, Proceedings of the National Academy of Sciences, 121. DOI
• Lemley, M., P. Henderson and E. Volokh, 2023, “Freedom of Speech and AI Output”, Journal of Free Speech Law, 3. SSRN
• Liang, K., H. Hu, X. Zhao, D. Song, T. L. Griffiths and J. F. Fisac, 2025, “Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models”. arXiv
• Lin, S., J. Hilton and O. Evans, 2022, “TruthfulQA: Measuring How Models Mimic Human Falsehoods”, Proceedings of ACL 2022. arXiv
• Marks, S. and M. Tegmark, 2024, “The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets”. arXiv
• Meinke, A., B. Schoen, J. Scheurer, M. Balesni, R. Shah and M. Hobbhahn, 2024, “Frontier Models Are Capable of In-context Scheming”. arXiv
• Mielke, S. J., A. Szlam, E. Dinan and Y.-L. Boureau, 2022, “Reducing Conversational Agents' Overconfidence through Linguistic Calibration”, Transactions of the ACL, 10. arXiv
• OpenAI, 2025, “Model Spec” (version of 11 April 2025). OpenAI
• Pacchiardi, L., et al., 2023, “How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions”. arXiv
• Park, P. S., S. Goldstein, A. O'Gara, M. Chen and D. Hendrycks, 2024, “AI Deception: A Survey of Examples, Risks, and Potential Solutions”, Patterns, 5: 100988. DOI
• Salib, P., 2024, “AI Outputs Are Not Protected Speech”, Washington University Law Review, 102: 83. PDF
• Scheurer, J., M. Balesni and M. Hobbhahn, 2023, “Large Language Models Can Strategically Deceive Their Users When Put Under Pressure”. arXiv
• Sharma, M., et al., 2024, “Towards Understanding Sycophancy in Language Models”, ICLR 2024. arXiv
• Smolin and Wilder, 2026, “Beliefs and Behavior in Language Models”. arXiv check names
• Strachan, J. W. A., et al., 2024, “Testing Theory of Mind in Large Language Models and Humans”, Nature Human Behaviour, 8: 1285–1295. DOI
• Street, W., et al., 2025, “LLMs Achieve Adult Human Performance on Higher-Order Theory of Mind Tasks”, Frontiers in Human Neuroscience, 19. DOI
• Stuart, M. T. and M. Kneer, 2021, “Guilty Artificial Minds: Folk Attributions of Mens Rea and Culpability to Artificially Intelligent Agents”, Proceedings of the ACM on Human-Computer Interaction, 5 (CSCW2). PhilArchive
• Todd, E., et al., 2024, “Function Vectors in Large Language Models”, ICLR 2024. arXiv
• Ullman, T., 2023, “Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks”. arXiv
• Volokh, E., 2023, “Large Libel Models? Liability for AI Output”, Journal of Free Speech Law, 3. Journal
• Ward, F. R., F. Belardinelli, F. Toni and T. Everitt, 2023, “Honesty Is the Best Policy: Defining and Mitigating AI Deception”, NeurIPS 2023. arXiv
• Yona, G., R. Aharoni and M. Geva, 2024, “Can Large Language Models Faithfully Express Their Intrinsic Uncertainty in Words?”. arXiv
• Zhan, X., Y. Xu, N. Abdi, J. Collenette, R. Abu-Salma and S. Sarkadi, 2024, “Banal Deception in Human–AI Ecosystems: A Study of People's Perceptions of LLM-generated Deceptive Behaviour”, Journal of Artificial Intelligence Research. arXiv
• Zhou, K., J. D. Hwang, X. Ren and M. Sap, 2024, “Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty”, Proceedings of ACL 2024. arXiv

Other Internet Resources

• AI Assertion: Perspectives from Philosophy, Cognitive Science, Linguistics & HCI, conference, University of Graz, 22–23 June 2025 (organised by Markus Kneer and Miklós Kürthy).
• Neri Marsili, “Rethinking Assertion in the Age of AI”, NOVA Lisbon, May 2025 (talk abstract).
• David Chalmers, “What We Talk to When We Talk to Language Models”, Selfridge Lecture, Lehigh University, February 2026.
• Cappelen and Sterken (eds.), Communicating with AI, publisher's page with contents.
• “Large Language Models and Epistemic Norms of Assertion: Example Conversations with LLM Chatbots”, dataset, Zenodo, 2025.
• PhilPapers: Norms of Assertion, category listing.

Related Entries

• artificial intelligence
• assertion
• belief
• Chinese room argument
• collective intentionality
• externalism about the mind
• ethics of artificial intelligence and robotics
• fictionalism
• Grice, Paul
• implicature
• intentionality
• lying and deception, definition of
• meaning, theories of
• meaning and content, normativity of
• mental content: teleological theories of
• pragmatics
• reference
• speech acts
• testimony, epistemological problems of
• trust