In the epistemology of testimony, a received view holds that testimony — assertion-based telling by a speaker — can itself generate knowledge for a recipient. Artificial systems now produce fluent propositional assertions (large language model outputs, predictions of fully automated epistemic systems), and it is contested whether such outputs can function as testimony in anything like the human testimonial sense, or only as non-testimonial evidence. The problem is open on two linked fronts, with a third available as an interpretive frame: (i) whether machine outputs can constitute testimony that yields knowledge for human recipients at all; (ii) specifically, whether the Kantian non-reductionist account of testimony — on which testimony is epistemically on a par with perception and presumptively to be trusted absent sincerity or competence defeaters — extends to machine testifiers or must be revised or restricted to humans; and (iii) optionally, how Neo-Kantian/transcendental conditions on knowledge, which treat knowledge as conditioned by conceptual and symbolic forms, bear on this assessment. The general question is actively contested and unresolved in 2019–2026 literature.

Background

Testimony, in the epistemology of testimony, is the acquisition of beliefs or knowledge from the assertions of others. The classical divide is between reductionism, on which testimonial beliefs must be independently corroborated by non-testimonial means such as perception, memory, or induction, and non-reductionism, on which testimony can by itself provide a basic warrant for justified belief. In the human case this debate is mature; the new question is whether its categories apply to machines.

Three strands of prior work frame the question. First, Humphreys (2009), in network epistemology — the study of how knowledge is produced and circulated in computer networks and distributed human–machine systems — compares epistemological issues arising in computer networks with standard features of social epistemology, provides a definition of knowledge for computational devices, and discusses nonconceptual content and testimony. On the testimony side, the account relied on here holds that only propositional, conceptual content counts as testimony in the human testimonial sense: when the computational side of a human–machine interface supplies only nonconceptual content (machine-level bits or property-specific demonstrations), those outputs cannot strictly be testimony, though they can function as epistemic evidence. Humphreys (2009) also explicitly left the fully automated scenario — in which networks acquire, model, and justify scientific knowledge without any human intervention, including what norms and criteria of acceptance obtain then — under-theorized.

Second, the Kantian non-reductionist account of testimony treats testimony about empirical matters as epistemically on a par with one's own experience, with a pro tanto presumption in favor of trusting a testifier absent relevant defeaters such as lack of sincerity or lack of competence. That account belongs to the Kant-on-testimony literature rather than being Mikalsen's own reconstruction: Gelfert (2006) is the study Mikalsen (2010) quotes and partly disputes, Mikalsen replacing Gelfert's gloss of the 'presumptive obligation' to trust with one grounded in the requirement to act only on universalizable maxims. Mikalsen's own thesis concerns Kant's idea of public reason — that it is 'neither at odds with deference to a certain kind of authority, nor with making judgements of character in rational deliberation', a view he says is corroborated by recent work on Kant's epistemology of testimony. Either way, this is a non-reductionist, presumptive-trust account for human testifiers; it says nothing about machines. The broader Neo-Kantian tradition (the Marburg school of Cohen, Natorp, and Cassirer, as reconstructed by Matherne 2015) adds a transcendental dimension: on the transcendental method, knowledge is not a copy of an independent reality but is conditioned by the conceptual and symbolic forms through which it is produced. Whether machine-generated propositional output can be a source of knowledge when knowledge is so conditioned is exactly the kind of question this tradition poses, and it has not been posed in it.

Third, recent work on AI as an epistemic agent has begun to contest the boundary. Ferrario, Facchini, and Termine (2024) argue, using accounts of expertise from virtue epistemology (the study of intellectual virtues and abilities as sources of epistemic worth) and Zagzebski's (2012) notion of an authority of belief — a source whose verdicts give one preemptive reasons to believe — that AI systems cannot be genuine epistemic experts because expertise requires a relation to understanding that AI lacks; at most, a human–AI hybrid agent, in which a human successfully appropriates the AI system, can be an expert or authority, warranting deference of belief but not understanding.

A rapidly growing 2019–2026 literature addresses machine testimony directly and divides over it. Freiman and Miller (2019) introduced quasi-testimony: technological devices can assert or testify under some conditions, despite relevant differences from human asserters. Freiman (2023) distinguished beliefs acquired from a conversational AI as technology-based beliefs, different in kind from both instrument-based and testimony-based beliefs, and Freiman (2024) argued that mainstream anthropocentric testimonial theories reject machine testifiers on three assumptions — that they lack intentions, cannot be normatively assessed, and cannot be objects of trust — and pressure-tested those assumptions. Williams and Bayne (2024) located chatbot output at the edge of assertion (the speech act of putting a proposition forward as true); Butlin and Viebahn (2025) defended the possibility of AI assertion against norm-based objections; He and Yang (2025) analyzed testimony by large language models directly. On the denial side, Sparrow and Flenady (2025) argue for a testimony gap: following Brandom's view that asserting is giving reasons in the space of reasons and so is an action for which one is responsible, machines — being neither moral agents nor reason-givers in this sense — cannot testify, and their outputs cannot be reasons for belief in the testimonial sense. DuVall (2026), in a Colorado State University master's thesis, concludes that LLMs can be credible experts only in limited circumstances, cannot be credible standard testifiers, and cannot contribute to group understanding. Hauswald (2025) examines whether AI systems can be artificial epistemic authorities, arguing that the identification problem (recognizing a reliable authority) and the deference problem (how much to rely) arise for AI in ways that parallel human authorities despite the absence of intentional belief transfer. Donner (2025) proposes artificially blended testimony for AI-generated documents, and Yue Weng Hong (2026) defends a non-anthropocentric notion of technological testimony. No position is settled; papers on all sides are appearing through 2026.

Problem Statement

Determine whether outputs of artificial systems — large language model assertions and the propositional outputs of fully automated epistemic systems, that is, outputs produced and justified without human intervention — can function as testimony that produces knowledge for human recipients, and in particular whether the Kantian non-reductionist account of testimony extends to machine testifiers. The knower whose testimonial knowledge is at issue is a human recipient; 'fully automated' qualifies the production side of the output, not machine-side knowledge. Two linked questions fall under this single objective. (1) The testimonial-status question: can machine-generated propositional output constitute testimony in the epistemologically load-bearing sense — rather than mere non-testimonial evidence, or the weaker categories proposed in the recent literature (quasi-testimony from technological devices, artificially blended testimony for AI-generated documents, or relocation of expertise to a human–AI hybrid agent) — such that a recipient can acquire testimonial knowledge from it? (2) The Kantian-extension question: does the non-reductionist, presumptive-trust account — testimony is epistemically on a par with perception and presumptively to be accepted absent defeaters — extend to machine testifiers; specifically, can its sincerity and competence defeater conditions be satisfied by, reinterpreted for, or shown inapplicable to artificial testifiers, or must the default entitlement be restricted to human testifiers? Neo-Kantian/Marburg conditions on knowledge — knowledge as conditioned by conceptual and symbolic forms, assessed by the transcendental method (Matherne 2015) — remain available as optional interpretive background that an answer may draw on, including for the fully automated production scenario; no verdict on whether they support, block, or reshape the extension is required. The question retains its natural generality across artificial testifier kinds (conversational LLMs, expert and predictive systems, instrument-like pipelines) and answer forms (affirmation, denial, or principled revision of the category of testimony); no restriction to a particular model class, domain, or answer form is imposed.

References

  1. Paul Humphreys (2009). Network Epistemology. Episteme 6(2), 221–229. DOI: 10.3366/e1742360009000653. URL: https://doi.org/10.3366/e1742360009000653
  2. Kjartan Koch Mikalsen (2010). Testimony and Kant's Idea of Public Reason. Res Publica 16(1), 23–40. DOI: 10.1007/s11158-009-9105-3. URL: https://doi.org/10.1007/s11158-009-9105-3
  3. Andrea Ferrario, Alessandro Facchini, Alberto Termine (2024). Experts or Authorities? The Strange Case of the Presumed Epistemic Superiority of Artificial Intelligence Systems. Minds and Machines 34(3). DOI: 10.1007/s11023-024-09681-1. URL: https://doi.org/10.1007/s11023-024-09681-1
  4. Samantha Matherne (2015). Marburg Neo-Kantianism as Philosophy of Culture. In The Philosophy of Ernst Cassirer: A Novel Assessment. De Gruyter. DOI: 10.1515/9783110421811-009. URL: https://doi.org/10.1515/9783110421811-009
  5. Linda Trinkaus Zagzebski (2012). The Authority of Testimony. In Epistemic Authority: A Theory of Trust, Authority, and Autonomy in Belief. Oxford University Press. DOI: 10.1093/acprof:oso/9780199936472.003.0007. URL: https://doi.org/10.1093/acprof:oso/9780199936472.003.0007
  6. Ori Freiman, Boaz Miller (2019). Can Artificial Entities Assert? In The Oxford Handbook of Assertion. Oxford University Press. DOI: 10.1093/oxfordhb/9780190675233.013.36. URL: https://doi.org/10.1093/oxfordhb/9780190675233.013.36
  7. Ori Freiman (2023). Analysis of Beliefs Acquired from a Conversational AI: Instruments-based Beliefs, Testimony-based Beliefs, and Technology-based Beliefs. Episteme 21(3), 1031–1047. DOI: 10.1017/epi.2023.12. URL: https://doi.org/10.1017/epi.2023.12
  8. Iwan Williams, Tim Bayne (2024). Chatting with bots: AI, speech acts, and the edge of assertion. Inquiry 69(6), 3059–3082. DOI: 10.1080/0020174X.2024.2434874. URL: https://doi.org/10.1080/0020174X.2024.2434874
  9. Ori Freiman (2024). AI-Testimony, Conversational AIs and Our Anthropocentric Theory of Testimony. Social Epistemology 38(4), 476–490. DOI: 10.1080/02691728.2024.2316622. URL: https://doi.org/10.1080/02691728.2024.2316622
  10. Robert Sparrow, Gene Flenady (2025). The Testimony Gap: Machines and Reasons. Minds and Machines 35(1). DOI: 10.1007/s11023-025-09712-5. URL: https://doi.org/10.1007/s11023-025-09712-5
  11. Jinhua He, Chen Yang (2025). Testimony by LLMs. AI & Society 40(8), 6201–6213. DOI: 10.1007/s00146-025-02366-y. URL: https://doi.org/10.1007/s00146-025-02366-y
  12. Patrick Butlin, Emanuel Viebahn (2025). AI Assertion. Ergo: An Open Access Journal of Philosophy 12. DOI: 10.3998/ergo.7960. URL: https://doi.org/10.3998/ergo.7960
  13. Sascha Donner (2025). What kind of documents can AI language model outputs be? The concept of artificially blended testimony. Journal of Documentation 81(3), 788–809. DOI: 10.1108/JD-11-2024-0271. URL: https://doi.org/10.1108/JD-11-2024-0271
  14. Rico Hauswald (2025). Artificial Epistemic Authorities. Social Epistemology 39(6), 716–725. DOI: 10.1080/02691728.2025.2449602. URL: https://doi.org/10.1080/02691728.2025.2449602
  15. Parker DuVall (2026). The epistemic limits of large language models: trust, testimony, and group understanding. Master's thesis, Colorado State University. DOI: 10.25675/3.027149. URL: https://doi.org/10.25675/3.027149
  16. Yue Weng Hong (2026). How is "AI testimony" possible? A "Non-Anthropocentric" Attempt to Defend "Technological Testimony". Dia-noesis: A Journal of Philosophy. DOI: 10.12681/dia.45790. URL: https://doi.org/10.12681/dia.45790
  17. Hyeongjoo Kim, Dieter Schönecker (eds.) (2022). Kant and Artificial Intelligence. De Gruyter. DOI: 10.1515/9783110706611. URL: https://doi.org/10.1515/9783110706611
  18. Axel Gelfert (2006). Kant on Testimony. British Journal for the History of Philosophy 14(4), 627–652. DOI: 10.1080/09608780600965226. URL: https://doi.org/10.1080/09608780600965226