I aim to investigate an increasingly important yet still underdeveloped topic in contemporary social epistemology, namely the epistemic status of beliefs formed on the basis of large language model outputs. As these models are rapidly integrated into social practices, the issues they raise seem to begin to implicate more fundamental philosophical questions.
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 physical signals. They may occupy a space in between, or perhaps a new kind of epistemic source not yet fully theorized.
Following recent discussions [@grzankowski2025], 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 both descriptive accuracy and 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 miss 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 [@bender2021]. 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 are in a way that allows Dutch-book vulnerabilities and intransitive preferences to arise quite naturally, perhaps even predictably [@GoldsteinManuscript-GOLLCN].
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 metaphysical undemanding beliefs), while remaining agnostic about heavier notions like consciousness or moral agency [@grzankowski2025]. 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 [@goldsteinlederman2025]. 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 [@levine2025resourcerationalcontractualismguide]). 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 this project, I hope to focus mainly on two strands, deflationism and Modest inflationism, as representing different strategies for interpreting LLM mentality.
This project aims to develop an epistemological framework for understanding LLM outputs. I plan to approach this goal through philosophical conceptual analysis that is informed by an interdisciplinary understanding of machine learning. The investigation is structured into four analytical stages. I first ask whether these outputs can plausibly be treated as a form of testimony. Then I will try to examine competing theories of justification and consider how far they can be applied to systems whose internal mechanisms do not mirror human mental kind. In the third stage, I aim to explore whether alternative models of group agency might offer a more suitable way of interpreting the behavior of LLMs. Finally, I apply the earlier findings to selected social and political contexts and try to determine whether any tentative conclusions might help illuminate the broader role of LLMs in public life.
The first stage focuses on asking whether LLM outputs can conceptually qualify as testimony. My method draws on a comparison of competing theories of testimony as they might apply to LLMs.
I plan to begin 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[@bender2021]. If one takes testimony to require intention, as in Coady, Graham, or Hinchman’s accounts, LLM outputs cannot qualify as testimony at all[@coady1992, @graham1997, @hinchman2005]. I will first consider this in relation to intention-based theories of testimony and try to see whether the absence of human-like intent blocks testimonial status or whether the barrier is less decisive than it first appears.
The intention worry will be less severe if we accept the Moderate Inflationist alternative. Some forms of testimony appear to depend on richer mental states such as consciousness, deliberation, or moral agency[@grzankowski2025]. For instance, words of comfort from our dearest friends seem to draw on the 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. Moderate inflationists, by contrast, argue that LLMs may possess lean mental states, such as F-belief, that carry low metaphysical demands. I aim to assess if attributing F-beliefs, might offer a middle ground that can grant fact-directed assertions (which might be metaphysically undemanding) as testimony and expressions requiring richer states (which would likely remain pseudo-testimony).
There is a way to bypass the intention requirement, for example, Lackey’s disjunctive account treats testimony as epistemically generative even when the speaker neither believes the proposition nor intends to inform the hearer, if the hearer has a valid reason to trust the speaker [@lackey2008]. In this case, I plan to explore what could be seen as a valid reason. It may be helpful 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(See some empirical evidence([@bisconti2025adversarialpoetryuniversalsingleturn])). So, I expect to argue that even on a disjunctive account, we have reasons to reject the first category, since fluency without minimal metaphysical grounding cannot meet even minimal testimonial conditions, though I am open to the possibility that the latter could support a modest and metaphysically undemanding form of testimony. This would allow us to exclude purely stochastic-parrot outputs while still identifying a limited yet epistemically meaningful testimonial pathway.
Even if LLM outputs are granted a provisional form of testimonial status, in the second stage, I will explore existing epistemological frameworks that might explain how such testimony could be justified (i.e., how one could have sufficient reason to believe that $p$). I do not assume in advance that any familiar framework will map neatly onto the case, and part of the task is to test how far each one can stretch before it becomes unstable.
I plan to begin by considering classical reductionism, 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, for example, perception, memory, or inference[@Fricker1987, @Adler1994]. 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 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(I assume that 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(Although, there is some normative & empirical work[@lazar2024frontieraiethicsanticipating]). 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. This is to say, we would have to be expert users in order to use LLMs. This issue can be framed in terms of 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, then they do not need the LLM’s testimony in the first place. If they lack this knowledge, then 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. This is not the intuition, because we think that in some cases, for example, when asking ChatGPT to write a piece of standard code or to answer a question of basic common knowledge, users can in fact acquire justified beliefs. To avoid these demanding requirements, one possible way I aim to explore is whether a move from Global to Local Reductionism allows us to salvage testimonial justification in ‘verifiable’ domains (like coding) while remaining skeptical in ‘open’ normative domains. However, given the limitations of this approach in ‘open’ normative domains, it is also necessary to evaluate the alternative, non-reductionist accounts.
I then turn to non-reductionist accounts that depend on the possibility of metaphysically undemanding F-beliefs, I wish to argue that non-reductionism may not be able to dismiss the epistemic burden in the same way as reductionism. For example, Reid’s account of our natural disposition to tell the truth includes a broader mechanism of belief-formation, one that does not depend on knowing the internal workings of other people’s minds[@reid1863]. If something analogous to a minimal truth-seeking disposition could be engineered in LLMs(See [@HeForthcoming-HETBLH]), 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. Because even if we could somehow engineer a stable truth-seeking tendency, it does not remove the requirement that the hearer remain sensitive or vigilant toward potential defeaters [@goldbergHenderson2006]. What I want to argue is that normal users not only lack the positive reasons mentioned above, but also lack any competence to recognize potential defeaters in LLMs’ output. Thus, even if the hearer can trust LLMs by default without any requirement for positive reasons, the sensitivity as a (negative) epistemic requirement would still rule out potential justification.
Once we recognize that ordinary users cannot meaningfully access either the positive or the negative epistemic reasons that might be tied to the internal processes behind an LLM’s outputs, it becomes increasingly difficult to say that the justification can rest entirely on the hearer’s own internal states. At least on my current understanding, the opacity of the model’s operation seems to push 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).
The third step involves examining anti-individualist reliabilism frameworks[@goldberg2010]. A crucial part of the analysis is the use of Barnett’s two sources problem [@barnett2015]. If we apply reliabilism to LLMs, this would mean that justification arises from the training data, the algorithms, the evaluation methods, and their collective dependability. However, if justification depends on the speaker’s belief-forming process, then the hearer may lack adequate internal reasons to distinguish between competing sources. Barnett argues that even if Sherlock is far more reliable than Clouseau, from the hearer’s perspective, they often lack the internal evidence needed to prefer Sherlock over Clouseau[@barnett2015]. If we nevertheless insist that the hearer ought to trust Sherlock(the belief that trust Sherlock is justified), Barnett thinks this borders on irrationality. When we bring this intuition into the LLMs’ context, the issue becomes more pronounced. 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.
My analysis suggests that 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, however, it seems that ordinary users lack any epistemic competence with respect to all of these demands. Against this background, I am inclined to think that 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.
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, as the relevant agent is the group, not the LLMs. 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, although it may soften their force. With respect to Reductionism in particular, 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. This trust concerns 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. I wish to argue that it seems reasonable to expect developers to offer fuller technical explanations and to move toward full open source. If these expectations are met, the public may gradually gain a clearer grasp of the strengths and limitations of these systems, and under such conditions, the epistemic burden imposed by Reductionism might become considerably more manageable.
Second, the core difficulty for Non-Reductionism is that users do not have the capacity to detect defeaters that arise from the model’s internal processes. Although defeaters for Non-Summativist group testimony are also hard to identify, 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. By relying on these mechanisms, users might bear a lighter burden when it comes to recognizing defective or unreliable models.
Third, the two sources problem (Barnett’s problem) may also become less severe. If model developers were to make their internal processes transparent, ideally through full open-sourcing, users would gradually gain access to reasons that allow them to distinguish stronger models from weaker ones. With such information in place, the rationality deficit involved in choosing between competing sources would be substantially reduced.
Finally, since the present 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. What I plan to address is that this highlights a necessary condition, namely 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 in a comprehensive way and submit themselves to robust oversight by public regulatory bodies. Only with this kind of institutional support can users reasonably accept such testimony.
Although I take the earlier philosophical questions to matter both for theory and for the development of LLM technologies. The broader social picture may be more at risk and consequential. If the time permits, I plan to explore how these concepts might affect our society.
The Gettier risk problem. This proposal mainly focuses on the problem of justification, but an important accompanying issue is the problem of knowledge. One question I want to explore is the kind of Gettier cases generated by LLMs. These refer to situations in which a subject has a justified true belief (JTB), but the generative mechanism of LLMs differs from what we traditionally consider as producing true beliefs. Borrowing Goldstein’s idea[@GoldsteinManuscript-GOLLCN], Goldstein explicitly 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).” Therefore, when a user acquires a true belief from an LLM, if the LLM generates that text on the basis of probability rather than factual or causal grounding, then the user’s belief-forming process contains an element of “epistemic luck.” What I want to explore is this, as suggested in discussions of reductionism, in certain specific domains (such as coding or mathematics), LLMs’ latent representation may actually achieve some level of isomorphism in specific domains. If such isomorphism could be demonstrated, then one might argue that this is not “luck” but a new kind of “non-causal truth-tracking,” which somehow avoid the Gettier objection.
The possibility of LLMs as a heuristic device. In epistemic democracy, individuals often lack the capacity to process complex political information[@brennan2016]. If future LLMs acquire stable functional representations, one might ask whether they could help citizens filter political facts without themselves acting as normative agents. The idea is that 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 failures remain significant problems.
Issues of distributive justice and collective labor. LLMs training depends heavily on publicly available texts and everyday linguistic contributions. Original creators typically lack meaningful awareness or control over their work. I think 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 emerging landscape is, I suspect, an important question for future work on digital-era distributive justice.
Although the earlier philosophical questions are important for both theory and the development of LLM technologies, I hope to conduct my serious thinking with a sense of realism and a broad perspective on how these theories affect our society (I believe that only by situating these discussions within a broader perspective can we fully express the spirit and significance of the humanities). My aim is to use these inquiries to clarify how modern knowledge-production mechanisms can coexist with social institutions and, in doing so, support and protect the living individuals and communities within society.
© 2026 CY. All rights reserved.
Stay updated
Occasional essays on LLM epistemology, alignment, and political philosophy. No spam.