2.1. Five accounts

Burge (Reference Burge1993) argues that hearers have an a priori entitlement to accept what they are told, absent countervailing reasons, because intelligible presentation-as-true is a prima facie sign of a rational source. Two features are easily misread. The entitlement is not keyed to the proximate medium: recordings, print, and machines can carry it, since what it licenses is the presumption of a rational source somewhere in the chain (Reference Burge1998: 17). And entitlement to accept is not yet transmission of knowledge: interlocution yields knowledge only if the extended body of warrant, which includes the resources possessed or indicated by the sources on which the recipient depends, contains enough to underwrite it (Reference Burge1998: 5–6). The load Burge places on the source is one of provenance and preservation.

Williamson (Reference Williamson2000: ch. 11) holds that assertion is constitutively governed by the knowledge norm: one must assert that p only if one knows that p. Whatever epistemic upgrade the hearer enjoys runs through the assertion’s status as an act regulated by that norm, so a producer whose assertions are not regulated by the aim of asserting only what one knows does not generate the same uptake. Goldberg (Reference Goldberg2010) holds that the process generating a testimonial belief includes the speaker’s own production, so the reliability relevant to the hearer’s warrant is in part the speaker’s; he sharpens this by contrasting reliance on an epistemic subject, assessable for rationality and responsibility, with reliance on a mechanism assessable for reliability alone, where the normative burden falls on the user (Reference Goldberg2012: 184, 188). Faulkner (Reference Faulkner2011) grounds testimonial warrant in assurance: the speaker invites affective trust, which presses a normative claim on its object, rather than the predictive trust one may place in any regular process, and thereby becomes answerable for the truth of what is said. A source so answerable is one for which the truth of what is asserted figures in why the assertion is made.

Lackey (Reference Lackey1999, Reference Lackey2008) has done most to weaken the speaker-side conditions. On her statement view, knowledge can be acquired from statements the speaker does not believe; the selfless-asserter cases, of which Stella the creationist biology teacher is the most famous, are designed to defeat the belief condition, and Section 4 grants that they do. But what Lackey puts in place of belief is itself a truth-connection: the speaker’s statement that p must be a reliable indication of p’s truth, and she must not possess undefeated defeaters. Reliable indication is a covariation notion. Even the most permissive account retains a substantive constraint between production and truth.

2.2. Fixing the strength

An argument of this shape is vulnerable to a charge of equivocation: a constraint weak enough to be common ground may be too weak to exclude anything, and a constraint strong enough to exclude may be nobody’s minimum. The remedy is to state the constraint once, at the strength intended, and to use it at that strength throughout.

Begin with a familiar case. A thermometer sits in a room held at constant temperature by a thermostat. It is broken; it would read the same whatever the temperature. Its readings are correct, and correct for a systematic, stable, explicable reason, yet no one thinks that posting them transmits knowledge of the temperature. The case fixes the target: the shared constraint must be strong enough to exclude the broken thermometer and weak enough to be genuinely common ground. Three candidate readings must be distinguished.

On a statistical reading, the required relation is non-accidental accuracy: a high rate of true assertions over the relevant range, stable, grounded, and explicable rather than a fluke. This is too weak. The broken thermometer’s accuracy is all of those things; what it lacks is any connection between each reading and the temperature that reading reports. If the source’s asserting p bears no explanatory relation to p’s truth, the hearer’s belief is true by a kind of luck that precludes knowledge, however impressive the aggregate record. That record can enter the hearer’s epistemic position, but it enters as inductive evidence about the source, and warrant so structured is the warrant of instrument use (Section 5). Grounded frequency reliability is non-accidental relative to the mechanism producing it while remaining accidental relative to the truth of the particular content asserted, and it is the second relation transmission requires.

On an internalist reading, the required relation is production governed by internal states that represent considerations as evidence, from the producer’s own perspective. This is too strong, and not common ground. Lackey’s cases show belief is not required; Williamson’s (Reference Williamson2006) externalism about understanding tells against building a first-personal grasp condition into the practice of assertion; and function-first accounts of assertoric norms (Kelp Reference Kelp2018) detach the norm’s content from the speaker’s psychology altogether. A TM that quietly reinstated an internalist grasp condition would beg the question against these positions and decide the machine case by stipulation.

The reading I intend, and the one I shall argue the five accounts jointly entail, is explanatory:

2.3. Entailment, and the objection from reliabilism

That each account entails TM can now be argued. Burge’s entitlement is underwritten by rational provenance and content preservation: a chain containing no source for which the truth of the content figures in its production contains no resources for the extended body of warrant to draw on. Williamson’s norm-regulated production inherits a truth-connection from the factive, modally robust state the norm invokes. Faulkner’s assurance entails answerability for truth, and answerability entails that truth figures in production. Lackey’s reliable indication is explicitly a covariation between statement and truth.

Goldberg deserves a paragraph rather than a clause, because an objection lives here. He is a process reliabilist, and process reliability is a frequency notion attaching to types. Had TM been stated as a condition of token sensitivity, his view would not have entailed it, and the convergence claim would have been four accounts and a remainder; the gap between conditional reliability of a type and sensitivity of a token is why tracking theories and reliabilism are distinct research programmes. Stated as it is, the entailment holds. Reliability is always reliability of a type, and the generality problem is not idle here: type the broken thermometer’s process as reading-a-thermometer-in-this-thermostatted-room, and it comes out conditionally reliable, so a reliabilism with no constraint on typing licenses the verdict that posting its readings transmits knowledge. Goldberg does not accept that verdict, and the line he draws is the one TM draws: he requires that the process whose reliability constitutes the hearer’s warrant be the operation of an epistemic subject, assessable for rationality and responsibility and not merely for reliability (Reference Goldberg2012: 184, 188). That is a demand that the relevant reliability be endogenous, a property of the source’s own epistemic operation rather than of a fortunate alignment between its outputs and the world.Footnote ¹

The accounts therefore differ over the grain at which the truth-connection must hold, not over whether it must: Burge, Williamson, Faulkner, and Lackey require it of the token assertion, Goldberg of the type the production instantiates. TM covers both, and nothing below turns on the difference, since the claim of Sections 3 and 4 is that current systems fail at every grain. The exogeneity alleged is not exogeneity of some particular output but of the production type: what connects such a process to the truth is the resemblance between conditions of use and conditions under which its encoded corpus regularities are truth-indicative.

2.4. The separability objection

A natural objection to the convergence claim runs as follows. The five accounts impose five different conditions, and the history of the field is a history of showing such conditions separable and individually unnecessary: Lackey detached belief, externalists detach grasp, functionalists detach speaker-centred norms. Why expect the residue to be stable?

The reply is that TM is not one of the conditions in the series, nor their conjunction, nor a disjunction of individually necessary components. It is the existential claim each account, and each successful weakening, entails: there must be some endogenous realisation of the explanatory truth-connection. Every published weakening replaces one realiser with another. Lackey installs reliable indication where belief was; Graham (Reference Graham2016) installs reliable production by properly functioning competence where Lackey’s package was; function-first theorists install a knowledge-generating function whose fulfilment condition, Section 3.2 argues, is TM again. Of an account that removes the connection itself the literature contains no instance, and the most developed cases for AI testimony (Faria Reference Faria2025; Freiman Reference Freiman2024b) argue that conversational systems satisfy the surviving conditions rather than that no truth-connection is needed. The disjunctive, open-ended form is not a weakness to be explained away; it is what the minimum of a contested literature has to look like.

3.1. Burge: entitlement and the conditions of Burge (Reference Burge1998)

Burge’s framework is often taken to exclude machine sources out of hand, on the ground that the entitlement is keyed to rational sources and machines are not rational. That is not his view, and the place to see this is the one text in which he addresses computational sources directly. In “Computer Proof, Apriori Knowledge, and Other Minds” (Reference Burge1998) he argues that a mathematician who accepts the Four-Colour Theorem on the strength of a computer-generated proof too long for any human to check can know, and that the warrant can remain a priori.

Three features of that discussion matter here. The default entitlement licensed by intelligible presentation-as-true is a warrant for presuming a rational source somewhere in the chain; one may even justifiably mistake a machine for a mind, and the presumption remains both warranted and, in the computer case, true (Reference Burge1998: 17). The machine contributes to knowledge because its operations mechanise the programmers’ rationality: its language and programme are “expressions of the mathematical rationality of the programmers” (Reference Burge1998: 8), and its output indicates a determinate justification, the derivation it executed, located in the extended body of warrant. And where the claimed result is difficult enough to raise reasonable doubt, the default entitlement requires supplementation (Reference Burge1998: 13), either by thinking through the source’s reasoning, incorporating fragments of it and appreciating its power “from the inside,” or by an induction over the source’s checked successes. Where the recipient has no insight into how the source reaches its results, one is “forced to treat the source as an object whose special rational reliability can be established only through induction on its past successes” (Reference Burge1998: 25).

Burge offers these not as necessary conditions on knowledge from any computational source but as an account of what makes the Four-Colour case work, and the argument below is correspondingly modest: the features doing the epistemic work there are absent in the LLM case, and the route his framework leaves open for LLMs is not the testimonial one.Footnote ²

First, provenance. There are rational sources behind an LLM in a thin sense: the authors of the training texts, the designers of the system. But Burge’s transmission condition is not existential in that thin way. The recipient’s knowledge depends on the chain containing, and the utterance indicating, justificational resources sufficient for the very content asserted.Footnote ³ The proof machine’s output stands in a determinate relation to a determinate ground, the derivation it carried out, which a mathematician can enter and check in part. An LLM’s output stands in no such relation to any particular upstream warrant. Data curation and preference training filter for quality, but imperfectly and at the level of sources rather than of the particular claim; nothing attaches to a given output the address of a ground.

Second, mechanised rationality. The proof programme’s transitions instantiate rules of inference the designers understood and certified as valid, which is why its running is an exercise, at one remove, of a warranted method. A trained network’s transitions were not designed and are not certified as truth-preserving. The designers’ rationality explains why the system is fluent and frequently accurate; it does not make any particular generative transition the execution of anyone’s warranted method. Duede and Davey (Reference Duede and Davey2025) press the point for the mathematical case: apriori knowledge is available from an opaque generative system only if a proof checker is attached, that is, only if the missing certification comes from outside. Where such a checker exists, Burge’s condition is met, and this argument does not apply; for open-domain assertion, none exists.

Third, supplementation. Incorporation requires that the recipient be able to think through the reasoning that produced the result. When a system is prompted to explain an answer already given, what it produces is a further output of the same generative process rather than a record of the process that produced the answer, and the displayed rationale can be demonstrably disconnected from the features actually driving the output (Turpin et al. Reference Turpin, Michael, Perez and Bowman2023). In systems that generate an extended reasoning trace before answering, the situation is better: the trace is causally upstream of the answer, and its role can be measured (Lanham et al. Reference Lanham, Chen, Radhakrishnan, Steiner, Denison, Hernandez, Li, Durmus, Hubinger, Kernion, Lukošiūtė, Nguyen, Cheng, Joseph, Schiefer, Rausch, Larson, McCandlish, Kundu, Kadavath, Yang, Henighan, Maxwell, Telleen-Lawton, Hume, Hatfield-Dodds, Kaplan, Brauner, Bowman and Perez2023). This is a real concession and marks a place where the Burgean route could open. But measured faithfulness is partial and varies with model and task, and causal upstreamness is not yet what Burge requires, which is that the recipient, by thinking the reasoning through, comes into contact with the grounds on which the assertion rests. A trace that shapes an answer without being governed by the evidential bearing of what it cites can be followed without thereby appreciating any justification.

What remains is Burge’s second route, induction over checked successes, and here a natural gloss must be corrected. It is tempting to say that induction over a track record is ipso facto empirical, so that this route delivers only empirical warrant. That is not Burge’s position: he argues at length that the third-person induction can itself be a priori, because its basis steps are apriori-warranted checkings of mathematical results the recipient thinks through (Reference Burge1998: 20–23). The correction leaves a better argument. Burge’s induction inherits its apriority from the apriority of the checked mathematics; where the subject matter is empirical, the safe paediatric dose or the date of a treaty, the basis steps are empirical and the induction is a track record in the ordinary sense. And Burge is explicit that the inductive route is what one is reduced to when one cannot incorporate the source’s reasoning, and that being so reduced means treating the source “as an object” (Reference Burge1998: 25). That is the instrument stance in Burge’s own vocabulary, and its structure, establishes a record under conditions of use and project it to the case at hand, is the structure of validating an instrument.

His footnote on Fermat’s Last Theorem completes the picture. A recipient can be default-entitled to accept a difficult claim, believe truly, and still fail to know, because the extended body of warrant lacks resources adequate to the claim’s difficulty (Reference Burge1998: 33n9). That is the situation of the chatbot user in Section 1: her proprietary entitlement may be intact and her belief true, and the transmission still fail, because there is nothing at the far end of the chain for the extended body to contain. The framework is hospitable to computational sources; the hospitality is conditional; the conditions are not met.

3.2. Williamson: grasp, the knowledge norm, and function-first alternatives

Two Williamsonian resources must be kept apart: the externalist account of understanding and the knowledge norm of assertion. On the first, to understand an expression is not to stand in some first-personal relation to its meaning but to participate, with sufficient fluency, in the practice of a linguistic community that uses it (Williamson Reference Williamson2006: 36–38). On that account, it is a serious question whether LLMs, trained on and causally coupled to the human practice at scale, understand what they produce, and nothing here answers it in the negative. The argument does not need that premise, and I do not assert it. Understanding is in any case compatible with every epistemic failing the knowledge norm marks: a human bullshitter understands his sentences perfectly.

The work is done by the second resource. On Williamson’s view, the hearer’s testimonial gain runs through the assertion’s status as an act regulated by the knowledge norm: the speaker’s production is shaped by the aim of asserting only what she knows. Because knowledge is factive and modally robust, production so regulated inherits a connection to truth and is one way of satisfying TM originally. The question about an LLM is therefore not whether it has attitudes but what regulates its production. On the standard description, production is regulated by conditional likelihood under the learned distribution, adjusted by preference optimisation against human ratings (Ouyang et al. Reference Ouyang, Wu, Jiang, Almeida, Wainwright, Mishkin, Zhang, Agarwal, Slama, Ray, Schulman, Hilton, Kelton, Miller, Simens, Askell, Welinder, Christiano, Leike and Lowe2022), which rewards what raters approve, that is, the appearance of norm-conformity. A process optimised to seem norm-conforming is not thereby regulated by the norm, for the same reason a student optimised to please graders is not thereby aiming at truth.

That description is now incomplete in a way that must be conceded rather than finessed. Contemporary post-training includes reinforcement learning with verifiable rewards, in which the reward model is replaced by a deterministic verifier such as a proof checker or a unit test (Lambert et al. Reference Lambert, Morrison, Pyatkin, Huang, Ivison, Brahman, Miranda, Liu, Dziri, Lyu, Gu, Malik, Graf, Hwang, Yang, Le Bras, Tafjord, Wilhelm, Soldaini, Smith, Wang, Dasigi and Hajishirzi2024). Where a verifier is in place, the training signal is correctness rather than approval, and production in that domain is regulated by a procedure whose function is to constrain output to truth. This looks like, and I think is, a case of original satisfaction.

The question is how far the concession reaches, and the position should be stated more carefully than a contrast between verifiable and unverifiable domains allows. Verification is not confined to formally decidable subject matters. Reward schemes keyed to whether an answer is supported by a retrieved authoritative document, or scored against a rubric applied by a further model, extend something in the neighbourhood of verification to history, medicine, and law, which are exactly the domains in which users treat these systems as informants. Two questions remain open, and the general claim turns on them rather than on a boundary drawn by decidability. The first is whether such schemes reward evidential bearing or the appearance of it: a rubric applied by a model trained on human approval may reproduce the correlation at issue, and a groundedness reward is satisfied by an answer matching a retrieved document irrespective of that document’s standing. The second is whether competence acquired under verification transfers to domains where nothing verifies. Neither is settled, and I do not claim otherwise. What I claim is that the burden on the transfer question falls on the party asserting the transfer, and that a defender of AI testimony who confined the thesis to domains where the reward signal is demonstrably keyed to truth would be making a case I do not here oppose.

A different move replaces the knowledge norm with a function-first one: assertion’s constitutive function is to generate knowledge in hearers, and one may assert that p only if the assertion is disposed to do so (Kelp Reference Kelp2018).Footnote ⁴ Since LLM outputs plausibly generate knowledge-like states in hearers, the thought runs, they satisfy the norm without speaker-side grasp. The reply is that function-first accounts relocate the truth-connection rather than removing it. An assertion fulfils a knowledge-generating function only if the hearer’s resulting belief is non-accidentally true, and hearer-side anti-luck cannot be secured by a token whose production is unconnected to the truth of its content: where frequency and truth diverge, a merely frequency-reliable practice generates true belief, not knowledge, and false belief with the same fluency. On function-first views, TM reappears as the norm’s fulfilment condition.

3.3. Goldberg: the locus and ground of reliability

Goldberg’s extended-process account requires that the process terminating in the hearer’s belief, which includes the speaker’s own production, be conditionally reliable with respect to the belief expressed. Two features of the LLM case block its application, and neither depends on any claim about current error rates.

The first is type individuation, where endogeneity does its work. The type an LLM’s production most naturally instantiates is emitting the continuation with high conditional likelihood under the learned distribution, given the prompt, and that type’s connection to truth is exogenous: it is truth-conducive when, and because, the conditions of use resemble conditions under which the corpus regularities it encodes are truth-indicative. Whether the production also instantiates a further, intrinsically truth-connected type is an empirical question assessed in Section 4.3. Until it is settled affirmatively, the reliability available to constitute the hearer’s warrant is exogenous, which is what Goldberg’s contrast between subjects and mechanisms excludes from the constituting role.

The second is projection. The reliability anyone can certify for such a system is measured under validation conditions and projects to the token case only on the assumption that the case is typical of what the measurements sampled, an assumption the system neither polices nor signals the failure of. Successive systems have become markedly more accurate, including in specialised domains, and nothing here denies it; the point is that improvement in the rate does not change what the rate is a rate of, or convert benchmark-relative reliability into the standing, subject-matter-directed property Goldberg’s warrant-constitution requires. Whether derivative assessment of designers and deployers closes the residual gap is the question Section 6 takes from Faria.

3.4. Faulkner: assurance and two kinds of trust

Faulkner’s account requires that the speaker be able to give assurance: to invite affective trust and take on answerability, to this hearer, for the truth of what is said. Artefacts are proper objects of predictive trust, and, on functional accounts, of trustworthiness understood as the disposition to fulfil functional obligations (Simion and Kelp Reference Simion and Kelp2023). Neither is the relation Faulkner makes warrant-grounding, and substituting them changes the explanans rather than satisfying it. The point is not that machines are beneath trust, but that Faulkner’s warrant is grounded in a relation whose existence entails the truth-connection. Section 6 asks whether the reductive and functional trust literatures supply an adequate substitute.

4.1. Stella and three rival readings

Stella is a high-school biology teacher and a devout creationist. She has mastered the evidence for evolutionary theory, teaches the consensus science competently and sincerely in the professional sense, and does not believe it; her students, on her say-so, come to believe and, Lackey claims, to know. I take the case at face value, but three rival readings must be set aside, because each would change what the case teaches.

The first holds that Stella believes what she asserts in some compartmentalised sense, so the belief condition survives (cf. Sosa Reference Sosa2007). I do not pursue it: it threatens to make the belief condition unfalsifiable, since any selfless asserter can be redescribed as a compartmentalised believer. I grant Lackey her case. The second, in the spirit of Fricker (Reference Fricker2006), holds that Stella is a conduit, the students’ warrant tracing to the textbook authors. Against this, Stella by stipulation selects, organises, corrects, and improvises; a student’s question gets an answer the textbook does not contain, and the warrant for that answer is Stella’s engagement with the evidential structure of the subject, not the book’s. But note what the objection could not win. If Fricker were right, Stella would satisfy TM derivatively, as a preserving link to sources satisfying it originally, and the case would still show that belief is unnecessary while a truth-connection is retained. The lesson survives either reading; the one I defend locates the satisfaction in Stella. The third reading replaces belief with reliable production by properly functioning competence (Graham Reference Graham2016), which is friendly: another way of satisfying TM originally.

4.2. Evidence-responsiveness without internalism

What, then, does Stella have that makes her case one of transmission? Apply the interventionist test. Hold the worldly evidence fixed and vary the proxy: introduce into her textbook an error of the kind her competence can detect. She does not assert it; she flags it, withdraws, consults further. Vary the evidence within the reach of what she can assess, and her assertions follow. Her production covaries with the evidential facts as far as her competence extends, and does so in virtue of her own cognition: the textbook does not screen the evidence off, because her competence can be brought to bear against it. That is original satisfaction of TM, stated without appeal to how things are for Stella from the inside. Her abstention from belief is an overlay, held for reasons that do not enter the production of her science teaching. Nothing requires that she represent considerations as evidence under a first-personal mode of presentation, so the constraint is not internalist and the concession to Lackey is not in tension with it. This is tracking reliability, in the vicinity of Nozick (Reference Nozick1981), as against the mere rate frequency reliability measures; original satisfaction just is the former realised by the source’s own production, and the latter, however high and however well grounded, realises nothing in TM by itself.

4.3. What current systems do: the state of the evidence

The question is whether the production of assertion-shaped outputs in current LLMs is tracking-reliable or only frequency-reliable. This is empirical, and flat assertion is inappropriate in either direction; here is the evidence as I read it.

The systems are trained, in the base stage, to maximise the conditional likelihood of tokens over a human text corpus, then adjusted by preference optimisation against human ratings (Ouyang et al. Reference Ouyang, Wu, Jiang, Almeida, Wainwright, Mishkin, Zhang, Agarwal, Slama, Ray, Schulman, Hilton, Kelton, Miller, Simens, Askell, Welinder, Christiano, Leike and Lowe2022) and, in verifiable domains, against deterministic verifiers (Lambert et al. Reference Lambert, Morrison, Pyatkin, Huang, Ivison, Brahman, Miranda, Liu, Dziri, Lyu, Gu, Malik, Graf, Hwang, Yang, Le Bras, Tafjord, Wilhelm, Soldaini, Smith, Wang, Dasigi and Hajishirzi2024). Retrieval-augmented generation conditions the model at inference on documents fetched from an external store (Lewis et al. Reference Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Küttler, Lewis, Yih, Rocktäschel, Riedel and Kiela2020), and deployed systems increasingly call external tools whose outputs they condition on. It is therefore not true that such systems cannot consult documents. The open question is what consultation amounts to: whether generation is governed by the evidential bearing of retrieved material, with sources weighed and conflicts adjudicated as such, or by likelihood given the augmented context.

The same question, one level down, is the question about internal representations. Probing studies find internal structure correlating with the truth of represented sentences (Azaria and Mitchell Reference Azaria and Mitchell2023; Burns et al. Reference Burns, Ye, Klein and Steinhardt2023); the structure is in some settings linear and decodable (Marks and Tegmark Reference Marks and Tegmark2024); and interventions along such directions at inference time can causally increase the truthfulness of outputs (Li et al. Reference Li, Patel, Viégas, Pfister and Wattenberg2023). This is genuine evidence that something in the vicinity of TM-relevant structure exists.

It is not yet evidence that TM is satisfied, and the friendly literature explains why. Herrmann and Levinstein (Reference Herrmann and Levinstein2025) propose standards an internal representation must meet to count as belief-like: accuracy, coherence, uniformity across domains, and use, where use requires that the representation govern the model’s outputs. The use standard is the TM-relevant one and the least established: probe-identified structure generalises poorly and is confounded with properties other than truth (Levinstein and Herrmann Reference Levinstein and Herrmann2025).

The gap between causal implication and governance carries much of the weight here. That a truth-direction can be intervened upon to make outputs more truthful shows that the direction is causally usable: it is a lever which, when pulled, moves production. It does not show that production is regulated by it when no one is pulling, which is what original satisfaction of TM and the use standard both require. A car’s heading can be changed by seizing the wheel, and that establishes nothing about what determines the route when the wheel is left alone. What would establish governance is evidence that, in unperturbed operation, variation in the represented truth value yields corresponding variation in what is asserted. Intervention studies are not designed to show this, and their authors do not claim that they do.

Behaviour points the same way. Models reproduce common human misconceptions, giving false answers that imitate widely attested text against the evidence (Lin et al. Reference Lin, Hilton and Evans2022), and their answers shift under user pressure without new evidence (Sharma et al. Reference Sharma, Tong, Korbak, Duvenaud, Askell, Bowman, Cheng, Durmus, Hatfield-Dodds, Johnston, Kravec, Maxwell, McCandlish, Ndousse, Rausch, Schiefer, Yan, Zhang and Perez2024), with expressions of uncertainty that are themselves outputs of the same generative process. The knowledge-conflict literature cuts both ways: when retrieved context contradicts a model’s parametric answer, models arbitrate rather than simply following the context, and the arbitration is sensitive to how coherent and specific the conflicting evidence is presented as being, and to how firmly the parametric answer is held (Longpre et al. Reference Longpre, Perisetla, Chen, Ramesh, DuBois and Singh2021; Xie et al. Reference Xie, Zhang, Chen, Lou and Su2024). That there is arbitration at all counts against the crudest versions of the present argument, and Section 8.2 takes it seriously. What has not been shown is that it is governed by evidential bearing rather than by presentational features correlating with it.

Every behavioural result cited is indexed to the model generations on which it was run, and successive systems have improved on all of them, so an argument resting on a particular error rate would be obsolete before publication.Footnote ⁵ This one does not: the claim concerns what governs production rather than how often production goes wrong, and the improvements are consistent with the ground of the reliability being unchanged. Nor does any of it require the deflationary description of these systems as mere mimics (Bender et al. Reference Bender, Gebru, McMillan-Major and Shmitchell2021), which this argument neither needs nor asserts. The honest form of the empirical claim is comparative and current: the best systematic description of what governs open-domain assertoric production remains conditional likelihood shaped by preference optimisation, and the hypothesis that truth-tracking representations govern it has not met, by its own proponents’ standards, the burden that would establish original satisfaction (cf. Shanahan Reference Shanahan2024).

The two conclusions distinguished in Section 1 carry different burdens. On the constitutive question, the burden is symmetric and the answer empirical; my claim is that the present balance of evidence favours the negative at the level that matters, the governance of production, and it is asserted with that force and no more. The practical question does not wait on it. A hearer who knows what these systems are has, in the facts rehearsed here, the standing grounds for doubt that suspend a default entitlement and trigger Burge’s demand for supplementation; and the supplementation available, validation of the record in the conditions of use, generates the use-based warrant of Section 5 rather than testimonial uptake. A hearer who does not know may retain prima facie entitlement, but transmission fails anyway, for want of anything at the far end for the extended body of warrant to contain, and that failure rests on the argument from provenance in Section 3.1 rather than on the contested question about governance. If interpretability research shows, for particular systems, that the use standard is met, those systems satisfy TM originally, and the verdict changes for them. That is not a concession extracted under pressure; it is what a structural constraint is for.

4.4. The argument stated

(P1) A source transmits testimonial knowledge that p only if the truth of p, or the evidential support for p, figures in the explanation of its producing the assertion that p, whether originally or derivatively, and only if the connection is endogenous to the source’s own production process (Section 2).

(P2) Stella satisfies TM without belief: originally on the reading defended here, derivatively on the conduit reading. Either way the belief condition falls, and the truth-connection stands (Sections 4.1 and 4.2).

(P3) On the best current evidence, deployed LLMs satisfy TM neither derivatively, generation being frequency-weighted synthesis rather than preservation of particular assertions with their grounds, nor originally, since open-domain production is governed by conditional likelihood shaped by preference optimisation, and the hypothesis of governance by truth-tracking representations has not met the standards its own proponents propose (Sections 3.1 and 4.3).

(C1) Therefore, on the best current evidence, current LLMs do not transmit testimonial knowledge, even on the most permissive account.

(C2) And independently of how the evidence bearing on P3 settles, hearers are not now positioned to accord LLM outputs testimonial uptake (Section 4.3).

8.1. The proxy-testimony objection

The strongest objection holds that an LLM is a proxy for the testimony of its corpus: when it outputs a well-attested truth, the assertion is routed through the human speakers whose warranted assertions trained it, so the chain of testimonial warrant runs through the model. Some versions add a principal, an institution for which the system asserts and which assumes responsibility for the output (Arora Reference Arora2024). This is the derivative route, and the routing metaphor fails at the aggregation step.

A preserving link transmits a particular assertion together with its provenance, which is why quotation, recording, and faithful relay carry warrant. Generation does not preserve particulars. The model’s parameters compress the corpus into distributional constraints and production selects continuations by likelihood under them, so a claim attested in many unreliable sources is favoured over one attested in few authoritative ones, and nothing weights by warrant or retains the address of a ground. The most careful friendly view here, on which LLM text has derivative content in something like the way library text does, itself stops short of treating outputs as preserved assertions, precisely because generation produces new text rather than conveying old (Lederman and Mahowald Reference Lederman and Mahowald2024). The proxy-with-a-principal version fails differently: a spokesperson relays a message her principal has authorised and would stand behind, whereas a deployer has not seen the output, does not believe it, and disclaims it in the terms of service. Responsibility for putting a system into the world is not authorship of its sentences.

Two boundary cases confirm the diagnosis. Where a model reproduces a memorised string verbatim, the output approximates preservation, but the system does not mark reproduction as reproduction, so quotation and synthesis are indistinguishable and even the degenerate case cannot serve as a preserving link. And the retreat to the corpus as a corporate testifier fails at one remove: an aggregate of texts weighted by frequency is not a source that stands in TM, since there is nothing whose relation to the truth of the particular output could do the explanatory work.

8.2. The collapse objection

The most pointed objection holds that Section 4’s contrast is too thin: Stella is reliable conditional on her textbook, the model conditional on its corpus, and both fail when their inputs fail. The obvious answer is an intervention: plant a falsehood in Stella’s textbook and she catches it; plant one widely in the corpus and the model reproduces it. That answer is not good enough, since the two interventions are not the same: one is local corruption of a single source, the other global corruption of the whole evidential environment. Equalise them and the easy verdict disappears in both directions. Under global corruption Stella fails too: a falsehood planted in every textbook, every teacher, and every journal she has read is one she will teach with full confidence, because her competence was built out of those sources. Under local corruption, the model does not simply capitulate: models arbitrate, and often reject the intruded claim, with the outcome depending on the coherence and specificity of the conflicting passage and on how firmly the parametric answer is held (Longpre et al. Reference Longpre, Perisetla, Chen, Ramesh, DuBois and Singh2021; Xie et al. Reference Xie, Zhang, Chen, Lou and Su2024). Both parties can override a corrupted proxy within some range and fail outside it, so the existence of override is not the difference.

The difference is what the override is responsive to, and it must be stated counterfactually rather than in terms of what Stella can say. Vary the evidential quality of the planted claim while holding fixed how well it is written, how confidently it is stated, and how authoritative its source looks: Stella’s verdict moves with the evidence, because the competence she brings to bear was built by engagement with the evidential structure of the subject rather than the presentational features of texts about it. Nothing turns on her being able to articulate her reasons; a teacher who could not say why a claim was unsupported would still decline to teach it, and articulacy is in any case something these systems supply readily. What the model’s arbitration is responsive to, so far as the evidence goes, are features of the conflict’s presentation that correlate with evidential bearing without being it: coherence, specificity, entrenchment of the parametric answer. Correlation is what one would expect, since these features are not accidentally related to good evidence. But a process that follows the correlate has its connection to the truth screened off by the proxy, in exactly the sense of Section 2.2.

This puts the disagreement where it belongs and makes it tractable, which I take to be a result rather than a concession. The claim I need is not that models never override attestation, since they do, but that their overriding is not governed by evidential bearing, and that claim is testable by the intervention TM specifies. Construct conflicting passages matched for coherence, specificity, and surface authority but differing in evidential quality, and see whether the arbitration moves with the evidence or the presentation. To my knowledge, the experiment has not been run in that form.Footnote ⁸ Until it is, the collapse objection and my reply are both promissory to that extent, and the case for the negative rests on the general characterisation of what governs production (Section 4.3). A dispute that reduces to a specifiable measurement is in better condition than one that does not, and the reduction is available here and not on accounts that turn on mentality.

8.3. The good-enough objection

Finally, it may be said that even if the classification is correct it is idle: in well-covered domains the LLM user ends up no worse off than the hearer of a competent informant. Two differences say otherwise. Parity holds only where distributional fit and evidential support coincide, and the coincidence is not signalled to the user. And the hearer’s ordinary resources for probing an informant, follow-up questions, requests for grounds, checks for consistency, have surface analogues in the interaction but not epistemic ones, since answers to follow-ups issue from the same process as the original answer and are not independent evidence about it.