The epistemic status of beliefs formed on the basis of large language model outputs is an increasingly important yet still underdeveloped topic in contemporary social epistemology. As these models are rapidly integrated into social practices, the issues they raise begin to implicate more fundamental philosophical questions.
The questions I keep returning to: Can the outputs of LLMs be regarded as testimony — and if so, what is the nature of that testimony? If they do constitute testimony, what could justify believing it? And might LLMs be better understood as a kind of non-summative group testimony? Further out, there are questions about whether LLM users are exposed to a kind of Gettier risk, about what LLM testimony does to our political life and democratic institutions, and about the distributive justice of it all.
This essay lays out how I currently think about these questions.
Where the debate stands
LLMs can generate flexible natural language that responds closely to context, so users often find themselves treating the model as if it were a testimonial agent. In such interactions, the model appears to take on the role of transmitting information and supporting belief formation as in the traditional testimony systems.
On the other hand, if we try to understand this phenomenon through traditional epistemological categories, the difficulties are that LLMs lack beliefs and understandings in the human sense, and they cannot bear normative responsibility. If we hold firmly to this view, it would seem that LLM outputs cannot ground justified belief. Yet this conclusion feels uneasy, since in many practical contexts these systems are highly reliable and can provide information that is accurate enough to guide action in medicine, law, finance, and other domains. If we dismiss their epistemic relevance solely because they lack certain metaphysical properties, we risk overlooking belief-forming practices that are already prevalent.
This suggests the need for a more fine-grained analysis. I am not certain where the eventual answer will lie, but I suspect we need to avoid both assimilating LLMs to human testifiers and reducing them to mere next-token predictors. They may occupy a space in between, or perhaps a new kind of epistemic source not yet fully theorized.
Following recent discussions (Grzankowski 2025), the debate over whether large language models have mental states can be divided into two broad positions. Inflationists argue that in some limited contexts, it is appropriate to attribute certain psychological states to LLMs. Deflationists, by contrast, maintain that such attributions lack metaphysical grounding.
Deflationists often rely on two strategies. The first is the Robustness Strategy, which emphasizes the model's sensitivity to small contextual shifts. Even semantically identical prompts may lead to noticeably different answers, suggesting that LLMs lack the stability required for attributing mental states. The second is the Etiological Strategy, which turns to the training mechanisms of next-token prediction and argues that such architectures lack communicative intentions or world-modeling capacities (Bender et al. 2021). From this viewpoint, LLM behavior looks more like statistical completion. However, human language typically has normative intentions.
Goldstein offers another explanation. He argues that the mechanism of next-token prediction lacks the features we normally associate with rational action. Since the model generates outputs by estimating the likelihood of a text continuation rather than by evaluating the expected value of any action, its architecture seems to make certain forms of incoherence almost inevitable. Under this reading, LLMs operate in a way that allows Dutch-book vulnerabilities and intransitive preferences to arise quite naturally, perhaps even predictably (Goldstein, manuscript).
Modest Inflationism offers a more cautious alternative. It allows the attribution of psychological kinds with relatively low metaphysical commitments, such as Fodorian belief (F-belief — leaner, metaphysically undemanding belief), while remaining agnostic about heavier notions like consciousness or moral agency (Grzankowski 2025). On this view, if internal structures in LLMs can play the functional roles associated with linking inputs to outputs, then a thin notion of belief might apply.
Goldstein and Lederman extend this approach through an interpretivist framework (Goldstein & Lederman 2025). They argue that mental attributions should depend on whether they yield sufficiently good predictions of behavior, assessed by empirical accuracy rather than by an interpreter's stance (for empirical research, see Levine et al. 2025). They also distinguish between model agents and instance agents, suggesting that psychological attributions apply only to the latter, since coherence emerges in the trajectory of a particular interaction and does not extend across instances.
In what follows I focus mainly on these two strands — deflationism and Modest Inflationism — as representing different strategies for interpreting LLM mentality. The argument moves in four steps: whether LLM outputs can plausibly be treated as a form of testimony at all; whether the familiar theories of justification stretch to cover them; whether group agency offers a better frame; and what any of this means for public life.
Is this testimony at all?
Start with the deflationist account. Deflationists often describe LLMs as next-token predictors. On this view, the model's only computational aim is to generate the statistically most likely next symbol. If we stay within this description, LLMs have no communicative intentions and no capacity to offer evidence to a hearer. An output such as "Beijing is the capital of China" is therefore treated as a probabilistically optimal string instead of an intentional assertion (Bender et al. 2021). If one takes testimony to require intention, as in Coady's, Graham's, or Hinchman's accounts (Coady 1992; Graham 1997; Hinchman 2005), LLM outputs cannot qualify as testimony at all. So the first question is whether the absence of human-like intent really blocks testimonial status, or whether the barrier is less decisive than it first appears.
The intention worry becomes less severe if we accept the Modest Inflationist alternative. Some forms of testimony appear to depend on richer mental states such as consciousness, deliberation, or moral agency (Grzankowski 2025). Words of comfort from our dearest friends, for instance, seem to draw on a richness of human spirit far beyond statistical prediction. Under a deflationist picture, LLMs can mimic the form of such expressions but lack the metaphysical grounding that gives them significance. Modest inflationists, by contrast, argue that LLMs may possess lean mental states, such as F-beliefs, that carry low metaphysical demands. Attributing F-beliefs might offer a middle ground: fact-directed assertions (which are metaphysically undemanding) could count as testimony, while expressions requiring richer states would remain pseudo-testimony.
There is also a way to bypass the intention requirement altogether. Lackey's disjunctive account treats testimony as epistemically generative even when the speaker neither believes the proposition nor intends to inform the hearer, so long as the hearer has a valid reason to trust the speaker (Lackey 2008). The question then becomes what counts as a valid reason. It helps to distinguish three kinds of outputs. The first consists of fluent strings with no metaphysical grounding — the classic "stochastic parrot" case. The second involves lean, belief-like functional states that can underwrite factual assertions. The third appears to require higher-order minds and phenomenal consciousness, such as deep emotional or moral utterances. Current LLMs seem to operate mostly between the first two; they still lack stable connections between latent representations and functional intentions (for some empirical evidence, see Bisconti et al. 2025). So even on a disjunctive account, I think we have reason to reject the first category — fluency without minimal metaphysical grounding cannot meet even minimal testimonial conditions — though I am open to the possibility that the second could support a modest and metaphysically undemanding form of testimony. This would let us exclude purely stochastic-parrot outputs while still identifying a limited yet epistemically meaningful testimonial pathway.
Paths to justification
Even granting LLM outputs a provisional form of testimonial status, the next question is how such testimony could ever be justified — how one could have sufficient reason to believe that p. I don't assume any familiar framework will map neatly onto the case; part of the exercise is testing how far each one can stretch before it becomes unstable.
Consider classical reductionism first. The concern is that reductionism places too much of an epistemic burden on hearers. Classical reductionism maintains that testimonial justification must rely on non-testimonial reasons the hearer already possesses — perception, memory, or inference (Fricker 1987; Adler 1994). On this view, users must have independent reasons to rationally justify the beliefs they form on the basis of an LLM's output. Yet reliable reasons to verify an LLM's outputs are epistemically demanding. Users need some way of checking that the model's output is reliable. In relatively closed domains such as coding, mathematics, or well-defined factual questions (assuming future models in these domains have stable connections between latent representations and functional intentions), we can fall back on stable benchmarks and shared standards of correctness, which provide users with a degree of independent support. But in more open-ended domains involving value disagreement, moral controversy, or social norms, the situation is much less clear (though there is some normative and empirical work — see Lazar 2024).
Even if we assume we can evaluate models in all possible domains, justifying our beliefs on the basis of an LLM's output would still require us to have sufficient knowledge of those domains. We would have to be expert users in order to use LLMs. The issue can be framed as a circularity: for a user to verify the reliability of an LLM on moral questions, they must already possess knowledge of moral theories. But if the user already has such knowledge, they don't need the LLM's testimony in the first place — and if they lack it, they cannot obtain any non-testimonial reasons to support the LLM's claims. As a result, normal users often lack a valid reason for assessing whether a given output is justifiable. That clashes with intuition: when asking ChatGPT to write a piece of standard code or answer a question of basic common knowledge, users surely can acquire justified beliefs. One possible way out is a move from Global to Local Reductionism — salvaging testimonial justification in "verifiable" domains (like coding) while remaining skeptical in "open" normative domains. But given the limits of this approach in open normative domains, the non-reductionist alternatives deserve a look.
Non-reductionist accounts depend on the possibility of metaphysically undemanding F-beliefs — but I don't think non-reductionism can dismiss the epistemic burden the way it hopes to. Reid's account of our natural disposition to tell the truth, for example, involves a broad mechanism of belief-formation that does not depend on knowing the internal workings of other people's minds (Reid 1863). If something analogous to a minimal truth-seeking disposition could be engineered into LLMs (see He, forthcoming), perhaps users would not need to understand training pipelines in detail, just as we do not understand the neural structure of human language. Yet current models do not seem to exhibit this kind of stability. And even if they eventually could, the defence feels incomplete: a stable truth-seeking tendency does not remove the requirement that the hearer remain sensitive or vigilant toward potential defeaters (Goldberg & Henderson 2006). My claim is that normal users not only lack the positive reasons discussed above — they also lack any competence to recognize potential defeaters in LLM outputs. So even if the hearer can trust LLMs by default without positive reasons, sensitivity as a (negative) epistemic requirement would still rule out justification.
Once we recognize that ordinary users cannot meaningfully access either the positive or the negative epistemic reasons tied to the internal processes behind an LLM's outputs, it becomes hard to say that justification can rest entirely on the hearer's own internal states. The opacity of the model's operation pushes us toward a picture where the hearer's mental perspective alone is not enough to ground the relevant epistemic support. (Local Reductionism can save expert users — not all users.)
Third, anti-individualist reliabilism (Goldberg 2010). A crucial part of the analysis here is Barnett's two sources problem (Barnett 2015). Applying reliabilism to LLMs would mean justification arises from the training data, the algorithms, the evaluation methods, and their collective dependability. But if justification depends on the speaker's belief-forming process, the hearer may lack adequate internal reasons to distinguish between competing sources. Barnett argues that even if Sherlock is far more reliable than Clouseau, the hearer often lacks the internal evidence needed to prefer Sherlock over Clouseau — and if we nevertheless insist the hearer ought to trust Sherlock, this borders on irrationality. Bring that intuition into the LLM context and the issue sharpens. The training and evaluation procedures of models are not transparent from the hearer's standpoint, and ordinary users typically do not understand them well enough to make meaningful distinctions. It becomes unclear how they could have sufficient reason to treat one model as justifiable and another not.
So all three major frameworks presuppose some level of epistemic competence on the part of the hearer. Reductionism assumes an ability to verify or evaluate supporting evidence. Non-reductionism assumes a sensitivity to potential defeaters. Reliabilism assumes an internal capacity to distinguish reliable from unreliable sources. In the case of LLMs, ordinary users seem to lack epistemic competence with respect to all of these demands. Against this background, I am inclined to think we probably have to abandon the attempt to ground justification in the internal mechanisms of the model's production, and instead ground it in elements that remain epistemically accessible to ordinary users — for example, LLMs as group agency.
LLMs as group testimony
There are several advantages to understanding LLMs as group agents. First, it allows us to avoid intention-based objections in debates about the nature of testimony: critics who argue that LLMs lack intentions can no longer dismiss their testimonial status on that basis, because the relevant agent is the group, not the LLM. Second, the justificatory grounds for group testimony are, in principle, more accessible to ordinary users.
If we proceed along this path, Summativism is clearly inadequate, since on the Summativist view a group's assertion that p amounts to the conjunction of its individual members' assertions that p. Given that no individual developer explicitly asserts every output of an LLM, Non-Summativism is the more appropriate path: the group's assertion that p is an emergent property of the group's testimonial act and does not require any individual member to believe p.
Adopting a Non-Summativist account does not remove the justification problems discussed earlier, but it may soften their force. With respect to Reductionism, the Non-Summativist model relocates the hearer's need for positive reasons. Instead of requiring detailed verification of the model's internal processes, the hearer is asked to maintain a more general trust in the developers and alignment teams — trust concerning the quality of datasets, the rigor of alignment procedures, and the standards used for safety evaluation. Admittedly, such trust is not easy to secure, especially given the novelty of LLMs and their growing social influence. But it seems reasonable to expect developers to offer fuller technical explanations and to move toward full open source. If those expectations were met, the public might gradually gain a clearer grasp of the strengths and limitations of these systems, and the epistemic burden imposed by Reductionism might become considerably more manageable.
The core difficulty for Non-Reductionism was that users cannot detect defeaters arising from the model's internal processes. Defeaters for Non-Summativist group testimony are also hard to identify — but human societies already maintain extensive regulatory frameworks for supervising corporations and other large organizations. It seems plausible that these external institutional checks could function as partial substitutes for the sensitivity that non-reductionist accounts usually demand from individual hearers. Relying on such mechanisms, users might bear a lighter burden in recognizing defective or unreliable models.
The two sources problem may also become less severe. If model developers made their internal processes transparent — ideally through full open-sourcing — users would gradually gain access to reasons for distinguishing stronger models from weaker ones, and the rationality deficit involved in choosing between competing sources would shrink substantially.
Finally, since this discussion concerns justification rather than knowledge, the question is whether a person has adequate reasons to believe that p, and whether the person may at some point be discharged from further epistemic responsibility. This highlights a necessary condition: the normative responsibilities borne by the model-training organization when it is understood as a Non-Summativist group agent. For their outputs to count as genuinely testimonial, and for users to be justified in accepting them, these organizations may need to open-source their systems comprehensively and submit to robust oversight by public regulatory bodies. Only with this kind of institutional support can users reasonably accept such testimony.
Open threads
The philosophical questions above matter both for theory and for the development of LLM technologies — but the broader social picture may be more at risk, and more consequential. Three threads I want to come back to:
The Gettier risk problem. This essay has mainly been about justification, but an important accompanying issue is knowledge. There may be a distinctive kind of Gettier case generated by LLMs: situations in which a subject has a justified true belief, but where the generative mechanism differs from what we traditionally consider as producing true beliefs. Goldstein points out that "selecting the most probable description given a prompt" is structurally completely different from "selecting an action based on its latent value (or truth)" (Goldstein, manuscript). So when a user acquires a true belief from an LLM that generated the text on the basis of probability rather than factual or causal grounding, the belief-forming process contains an element of epistemic luck. But as the reductionism discussion suggested, in certain domains (coding, mathematics) the model's latent representations may achieve some level of isomorphism with the domain. If such isomorphism could be demonstrated, one might argue this is not "luck" but a new kind of non-causal truth-tracking — which would somehow avoid the Gettier objection.
LLMs as heuristic devices for democracy. In epistemic democracy, individuals often lack the capacity to process complex political information (Brennan 2016). If future LLMs acquire stable functional representations, could they help citizens filter political facts without themselves acting as normative agents? A model with transparent reasoning paths might serve as a heuristic device, supplementing public deliberation without requiring its own political rationality. Still, risks of inherited bias, misalignment, and governance failure remain significant.
Distributive justice and collective labor. LLM training depends heavily on publicly available texts and everyday linguistic contributions, and the original creators typically have no meaningful awareness or control over that use. This probably constitutes an unacknowledged form of collective labor. It remains unclear whether new institutions for data ownership or benefit-sharing should be developed, but the concentration of economic value in model companies seems increasingly difficult to justify. How fairness should be addressed in this landscape is, I suspect, an important question for digital-era distributive justice.
References: Adler (1994); Barnett (2015); Bender et al. (2021); Bisconti et al. (2025); Brennan (2016); Coady (1992); Fricker (1987); Goldberg (2010); Goldberg & Henderson (2006); Goldstein (manuscript); Goldstein & Lederman (2025); Graham (1997); Grzankowski (2025); He (forthcoming); Hinchman (2005); Lackey (2008); Lazar (2024); Levine et al. (2025); Reid (1863).