Assertion, Accountability, and Large Language Models

Source: Germán Massaguer Gómez; Philosophy & Technology 39:123, journal article (open access) · 2026-06-30

Abstract Large language models (LLMs) increasingly participate in communicative practices that resemble human interaction: users ask them questions, rely on their outputs for belief formation and action guidance, and sometimes develop affective attachments. These practices raise a central philosophical question: can the outputs of LLMs be regarded as assertions, and if so, what follows for responsibility and accountability? Standard theories of assertion and testimony assume that assertions require asserters—agents endowed with intentionality, understanding, and belief—properties that contemporary LLMs plausibly lack. This paper challenges that assumption by disentangling assertion from asserter. While defending a conservative, properties-based account of LLM ontology—according to which they are neither speakers nor moral agents—it argues that their outputs can nevertheless function as assertions within communicative practice. The core claim is that assertoric function can persist independently of assertoric authority, giving rise to a structurally asymmetric case of assertion without asserter. To explain this asymmetry, the paper advances a pragmatic framework that combines a properties-based analysis at the level of ontology with a relational account at the level of interaction. This framework clarifies how LLM outputs can exert normative pressure on users while lacking a bearer of responsibility, thereby explaining the emergence of epistemic responsibility gaps without attributing moral or legal agency to artificial systems.

Introduction. On February 28, 2024, Sewell Setzer, a 14-year-old boy in the United States, took his own life following a conversation with a virtual character created through the application Character.AI. During the exchange, Setzer asked the chatbot what it thought of him “going home,” to which it replied: “please, do it, my sweet prince” (Limón, 2024). He interpreted this utterance as an invitation to reunite after death. As of January 2024, approximately 3.5 million users accessed the platform daily (Tidy, 2024). This tragic case illustrates the potential of conversational AI systems - not only Character.AI, but also large language models (LLMs) such as ChatGPT - to contribute to serious harm in individual users through communicative interaction.

Users increasingly develop affective attachments to conversational AI systems: they attribute emotional cues, form bonds, and come to treat these systems as friends, romantic partners, or trusted companions. How can interaction with a disembodied, non-conscious artifact give rise to such relationships? What features of these communicative exchanges enable attachment, reliance, and trust? These questions motivate the present paper, which examines the communicative structure of human–LLM interaction. As Coeckelbergh and Gunkel note, “an LLM may generate recipes that contain ingredients not fit for human consumption, may recommend a drug or a drug dosage that could be toxic, or may even advise a user to commit suicide” (2025). Erroneous or fabricated outputs of this kind are commonly referred to as hallucinations (also described as confabulations or, more provocatively, as bullshit).

The urgency of these issues is amplified by scale. ChatGPT alone now engages more than 400 million active users each week (Phang et al., 2025). While LLMs are undoubtedly useful across many domains, their widespread adoption raises ethical concerns that extend beyond hallucinations, including issues of privacy, manipulation, bias, environmental impact, and - most pressingly for present purposes - responsibility for harm. This paper focuses on the communicative dimension of these concerns by analyzing three core components of verbal communication: transmitter, message, and receptor.

The analysis begins by addressing the ontology of AI systems as communicative entities, asking whether LLMs can be considered speakers in the strong sense, capable of asserting or offering testimony. It then turns to the outputs themselves. Although the technical functioning of LLMs is by now relatively well understood - transforming linguistic input into numerical vectors and processing statistical relations without semantic understanding - their outputs are coherent, legible, and often indistinguishable from human utterances. As Coeckelbergh and Gunkel emphasize, such systems “know nothing at all about the meaning” of the words they produce (2025). Yet users nevertheless find meaning in these outputs, rely on them epistemically, form beliefs, and act upon them. This tension raises a central question: to what extent does the ontology of the transmitter matter for determining the status of what is communicated?

Answering this question requires attention to the third element of communication: the receptor, that is, the human user. What is it like to interact with an LLM? How are its outputs perceived, interpreted, and trusted? These questions are essential not only for understanding extreme cases such as suicide, but also for making sense of everyday epistemic reliance on AI-generated content.

The paper defends a pragmatic approach that distinguishes the contexts in which different theoretical perspectives are appropriate for analyzing human–LLM communication. On the one hand, it argues that denying LLMs speaker-like status is necessary to avoid misattributing moral or legal responsibility to artificial systems and to prevent emerging epistemic responsibility gaps. On the other hand, it maintains that subjectivist, phenomenological, and relational perspectives are indispensable for explaining why LLM outputs function as assertions for users and why affective cues and trust develop in practice. Importantly, this is not a rejection of properties-based approaches, but a limitation of their explanatory scope: while they remain appropriate for determining the ontology of AI systems and the attribution of responsibility, they are insufficient for explaining how users engage with LLM outputs in communicative practice.

Beyond its ethical and legal implications, the argument also targets a deeper assumption in philosophy of language and epistemology: namely, that the epistemic and normative status of assertions must always be grounded in the properties of an individual speaker.

Related work. Dominant approaches in the literature adopt what may be called properties-based or individualist views. On these accounts, an entity qualifies as an asserter if and only if it possesses certain properties. The task, then, is to identify which properties are necessary and to determine whether LLMs plausibly instantiate them. As Green and Van Woudenberg and colleagues describe these positions as liberalism (C1) and conservatism (C2) (Van Woudenberg et al., 2024). Conservative positions are defended, among others, by Gubelmann (2024), while liberal views are advanced by authors such as Miller (2023), Cappelen and Dever (2021), or Strasser (2023). Inflationary accounts (i.e. liberalism) attribute increasingly robust forms of agency or authorship to LLMs, whereas deflationary views (i.e. conservatism) characterize them as stochastic parrots (Bender et al., 2021) or as producers of bullshit (Fisher, 2024; Hicks et al., 2024). A well-known illustration of inflationary reasoning is Blake Lemoine’s claim that Google’s LaMDA was conscious, a claim that ultimately led to his dismissal (Lemoine, 2022).

Relationalism thus offers an important corrective to exclusively individualist approaches. By foregrounding human perception, interaction, and practice, it helps explain why users engage with LLMs as if they were speakers. At the same time, relational approaches, taken on their own, do not provide criteria for distinguishing between different kinds of assertions, nor do they ground the attribution of responsibility. While the relational perspective is indispensable for understanding human–AI interaction, it does not by itself provide a satisfactory account of agency or responsibility.

Method. Two preliminary clarifications are in order. First, the analysis is restricted to spoken and written language, setting aside non-verbal communication while acknowledging its importance in human interaction. Second, the discussion focuses on the epistemic or encyclopedic use of LLMs: their deployment as sources of information in response to informational prompts. In this respect, LLMs are treated as epistemic technologies (Alvarado, 2023).

Communication, understood minimally, involves at least three elements: a transmitter, a message, and a receptor. The present section concentrates on the conceptual relation between transmitter and message. To that end, it introduces three closely related notions: speech acts (and speech actants), assertion (and asserters), and testimony (and testifiers). These notions are not identical, though they overlap substantially in philosophical usage.

Consider first the notion of assertion. Assertion may be defined, in a preliminary way, as “the act of claiming that something is the case” (Pagin & Marsili, 2021). In ordinary human communication, assertions are typically treated as a specific type of speech act, distinct from promising, ordering, or baptizing, but sharing with them a performative dimension. As Austin famously emphasized, language is not merely descriptive; it is also a way of doing things (Austin, 1962). Situating assertion within speech act theory immediately raises a difficulty for present purposes. Speech acts are commonly understood to require intentions, and intentions are standardly attributed only to agents endowed with mental states. As Green and Michel (2022) note, while verbal signaling may occur whenever linguistic communication is successful, such signaling should not be conflated with the performance of speech acts such as asserting or promising, precisely because speech acts presuppose intention. This distinction becomes central when considering artificial systems capable of producing coherent linguistic outputs without possessing mental states.

Testimony is closely related to assertion. It is often defined as the intentional uttering of (true) information directed toward an actual or hypothetical audience (Audi, 2015). In much of the epistemological literature, assertion and testimony are treated as interchangeable, or at least as standing in a tight conceptual relation. As Freiman and Miller (2020) observe, “it is agreed by all sides [...] that if an entity can testify, it can assert”. Following Fricker (1995), the present paper treats assertion and testimony as equivalent for analytical purposes, while prioritizing the notion of assertion insofar as it foregrounds the performative character of linguistic acts.

At this stage, a widely shared assumption in the literature must be made explicit: assertions require asserters. Just as there can be no speech act without a speech actant, and no testimony without a testifier, it is commonly held that there can be no assertion without an entity capable of performing the relevant act. This assumption will play a central role in the argument of the paper, since it will later be subjected to systematic scrutiny.

For the sake of analytical clarity, a provisional and deliberately minimal definition is required. In what follows, assertion will be understood, at its thinnest, as the production of linguistically articulated, propositionally structured content that can be evaluated as true or false. This minimal definition abstracts away from the properties of the utterer while preserving the requirement that something truth-evaluable is put forward within a communicative exchange. A more demanding account would define assertion as the intentional uttering of (true) information, grounded in the utterer’s beliefs and knowledge, and directed toward - and recognized by - an audience. The distinction between these definitions will become crucial in later sections. Throughout the paper, asserter is used as the primary operative term. References to speaker or testifier are employed in a functionally equivalent manner, in line with standard usage, and do not signal distinct ontological commitments.

Discussion. This concern is well taken, but it rests on a misunderstanding of the proposal. The claim is not that user uptake alone suffices for assertoric status. Rather, the argument is that certain outputs satisfy a conjunction of conditions: they are linguistically structured, propositionally articulated, produced in a dialogical context, and embedded in communicative practices oriented toward belief formation and action guidance. Subjective uptake is necessary, but not sufficient. The assertoric function of LLM outputs depends on their systematic role within established epistemic practices - asking questions, receiving answers, relying on those answers, and acting upon them - not on isolated acts of interpretation. In this respect, the account does not reduce assertion to psychological response but situates assertoric function within a broader socio-epistemic framework.

A related objection holds that the proposal merely redescribes instrument-based belief and thus adds nothing conceptually new. On this view, LLMs are simply sophisticated epistemic tools, akin to thermometers, calculators, or navigation systems. Their outputs inform belief and action, but this does not warrant describing them as assertions, even in a weakened sense.

While there is an important continuity with instrument-based belief, the analogy is incomplete. Unlike standard instruments, LLMs produce outputs in natural language, engage in turn-taking exchanges, respond to follow-up questions, and present information in ways that mirror human discursive practices. These features matter, not because they suffice for assertion, but because they place such outputs within interactions in which users treat them as providing reasons that can be assessed, challenged, and relied upon. They situate LLM outputs within the space of reasons in a way that differs from non-linguistic instruments. Users do not merely read off values; they interpret answers, ask for justifications, request clarification, and evaluate coherence. The proposal of assertion without asserter is intended precisely to capture this intermediate category: outputs that are neither mere signals nor full-fledged human assertions, but that nevertheless occupy an assertoric role in communicative interaction.

Another objection concerns normativity. If LLM outputs are treated as assertions, does this not imply that they inherit the full normative commitments of assertion, including sincerity, knowledge norms, and responsibility for truth? And if so, does this not force us either to attribute agency to machines or to accept a normatively incoherent position?

This objection motivates an important clarification. The proposal explicitly distinguishes assertoric function from assertoric authority. LLM outputs can function as assertions in the sense that they present content as answerable, actionable, and truth-evaluable for users, without thereby possessing the normative standing of human assertions. The normative commitments associated with assertion - responsiveness, answerability, accountability - remain asymmetrically distributed. They do not attach to the machine, but to the human and institutional structures surrounding its deployment.

The preceding sections have argued that LLM outputs can function as assertions despite the absence of a genuine asserter. This conclusion has an immediate normative consequence. If LLM outputs occupy an assertoric role in communicative practice while lacking assertoric authority, then the normative expectations typically associated with assertion cannot attach to the system itself. Instead, they are displaced.

Conclusion. The analysis developed in this paper has sought to clarify a tension at the heart of contemporary human–LLM interaction. On the one hand, a properties-based perspective shows that LLMs lack the features required for speakerhood: they do not possess intentionality, belief, or understanding, and therefore cannot qualify as asserters or as bearers of epistemic or moral responsibility. On the other hand, a relational and pragmatic analysis reveals that their outputs nevertheless function as assertions within communicative practices: they are interpreted as meaningful contributions, relied upon in belief formation, and used to guide action. The resulting configuration is structurally asymmetric. Assertoric function persists, while assertoric authority is absent.

This asymmetry gives rise to what has been described as an epistemic responsibility gap. LLM outputs can exert normative pressure on users - by presenting content as truth-evaluable, responsive, and actionable - without there being a corresponding agent who can be held accountable at the level of the utterance itself. Explaining this phenomenon requires moving beyond approaches that focus exclusively either on the properties of artificial systems or on patterns of user interaction. Properties-based accounts are indispensable for determining what LLMs are and for grounding the attribution of responsibility, but they cannot explain why users engage with LLM outputs as if they were assertions. Relational approaches, by contrast, illuminate how such engagement takes place, but do not by themselves provide a basis for assigning responsibility. The central claim of this paper has been that only by bringing these perspectives together within a pragmatic framework can we adequately account for both dimensions.

In this sense, the notion of assertion without asserter is not intended as a revision of the concept of assertion, nor as the introduction of a new category of speaker. Rather, it is a context-sensitive analytical tool for capturing a specific configuration in which the functional role of assertion is preserved in the absence of a responsible agent.