In 'Network Epistemology' (Episteme, 2009), Paul Humphreys initiated an epistemology for networked computational systems that produce scientific knowledge: he proposed a model-based definition of knowledge for computational devices, argued that nonconceptual machine outputs supply evidence rather than testimony, and defended a broadly reductionist stance on accepting instrument and simulation outputs. The paper explicitly left five issues open, and together they define the problem posed here: the epistemic norms, responsibility structures, and model-acceptance criteria for fully automated networks with no human agents; operationalization of the truth condition of the model-based definition under approximation and idealization (selective realism, measurable approximate truth); principled criteria for individuating knowing machines and networks; design principles for converting nonconceptual machine representations into reliable, transparent propositional content for human users; and community norms for epistemically assessing large, opaque simulations.

Background

How should epistemology, the philosophical study of knowledge and justification, handle scientific work that is done by machines? Since the late twentieth century, large parts of scientific data acquisition and analysis have been automated: robotic telescopes schedule and execute observations, particle-physics detectors record collisions and trigger on interesting events, genomic pipelines assemble and annotate sequences, and large simulations produce results that no human could recompute by hand. The branch of epistemology that studies knowledge as a social achievement, social epistemology, whose central topics include testimony (knowledge gained from another agent's say-so) and trust, was built for human agents. Paul Humphreys' article 'Network Epistemology' (Episteme, 2009) starts from the observation that this framework is anthropocentric, and asks what happens to its core concepts when the epistemic agents are computer networks.

Humphreys' paper makes four principal moves. First, it offers a model-based definition of knowledge for computational devices: a device knows about a system S roughly when it possesses a model M of S that is true (accurately represents the relevant aspects of S) and is evidentially supported. Second, it argues that scientific instruments and computers can produce and manipulate nonconceptual content, that is, representational content that does not consist of concepts or propositions (analogous to perceptual content), as when a detector registers a property of the world without representing it propositionally. Third, because testimony in the human sense requires propositional, conceptual content, machine outputs that are nonconceptual function as evidence rather than testimony; accepting them therefore cannot follow the standard testimonial route. Fourth, for hybrid human-machine inquiry he endorses a broadly reductionist stance he calls local reductionism: relying on an instrument's or simulation's output is justified by positive reasons, such as knowledge of how the device works, calibration, and understanding of error sources, rather than by default trust. He also proposes a functional criterion for individuating epistemic agents in networks, based on when models and knowledge bases can be transferred intact between nodes, and notes that computers can run many models in parallel, relaxing human-style demands for global consistency of an agent's beliefs.

The paper itself closes with an explicit list of unresolved issues, which constitute the problem posed here. First, the fully automated scenario, in which networks acquire, model, and justify scientific knowledge without any human intervention, is under-theorized: what norms, responsibility structures, and criteria of model acceptance obtain when no human performs calibration, sets domains, or adjudicates defeaters (considerations that would defeat a knowledge claim)? Second, the truth component of the model-based definition needs analysis when models contain approximations and idealizations: selective realism, the realist view that only particular components of successful models are true representations, must be operationalized, and approximate truth measured, if the definition is to apply to real models. Third, the individuation criterion is incomplete: the paper assumes a functional, resource-based cutoff for when a single machine can stand in for a network, but the theoretical limits and concrete metrics (memory, processing throughput, latency, representational completeness) remain to be formulated. Fourth, the interface problem: how can pervasive nonconceptual machine representations be converted into propositional content reliably and transparently for human users, with appropriate epistemic safeguards? Fifth, the epistemic assessment of large, complex simulations is impeded by epistemic opacity, Humphreys' term for the inaccessibility of all relevant steps of a computational process, and by unpublished implementation details; community norms of disclosure, reproducibility, and domain-specific reliability assessment are needed so that local reductionism can be practically implemented.

Since 2009 these strands have been taken up in partly separate literatures: on whether artificial systems can assert or testify; on responsibility for autonomous systems; on trust in computer simulations and machine learning; on extended and collective knowledge; and, most recently, on the epistemology of autonomous 'AI scientists'. The problem below asks for the completion of Humphreys' framework in the light of that literature. Note that the phrase 'network epistemology' has also come to name a different research program, on communication networks of human scientists; the problem here is the one Humphreys posed about machines.

Problem Statement

Complete the epistemology of networked, automated science that Humphreys (2009) initiated, by resolving, as one coherent framework for when and how machine-generated states count as scientific knowledge that humans and other machines may accept, the five issues his paper explicitly left open:

  1. Fully automated scenario. Specify the epistemic norms, responsibility structures, and criteria for accepting models produced by networks that acquire, model, and justify scientific knowledge without any human intervention. In particular: who or what bears epistemic responsibility for machine-generated knowledge claims, and what plays the roles that human calibration, domain-setting, and adjudication of defeaters play in hybrid inquiry?

  2. Truth under idealization. Operationalize the truth condition of the model-based definition of computational knowledge for models that employ approximations and idealizations: give principled criteria for which components of such models count as true representations of which aspects of their target systems, and specify how approximate truth is to be assessed, including any measurable notion of it.

  3. Individuation of knowing subjects. Formulate the theoretical limits and concrete criteria, including resource-based metrics such as memory, processing throughput, latency, and representational completeness, that determine when a single machine can stand in for a network as the epistemic subject, refining or replacing the source's functional transferability criterion.

  4. The interface problem. Provide design principles and epistemic safeguards for converting nonconceptual machine representations into propositional content that human users can adopt reliably and transparently, respecting the source's distinction between testimonial and merely evidential output.

  5. Assessment of large, opaque simulations. Establish the grounds and community norms, concerning disclosure, reproducibility, and domain-specific reliability assessment, under which acceptance of results from large, complex simulations whose implementation details cannot all be published can be justified, so that broad-based local reductionism can be practically implemented.

These five issues are to be resolved together, so that the results jointly amount to a working epistemology of machine-generated scientific knowledge; the source's model-based definition of computational knowledge and its local-reductionist stance may be revised where the argument requires, and local reductionism need not be retained as the framework that issue 5 makes practically implementable.

References

  1. Humphreys, Paul (2009). 'Network Epistemology'. Episteme 6(2), pp. 221-229. DOI: 10.3366/e1742360009000653. https://doi.org/10.3366/e1742360009000653
  2. King, Ross D.; Rowland, Jem; Oliver, Stephen G.; Young, Michael; Aubrey, Wayne; Byrne, Emma; Liakata, Maria; Markham, Magdalena; Pir, Pinar; Soldatova, Larisa N.; Sparkes, Andrew; Whelan, Kenneth E.; Clare, Amanda (2009). 'The Automation of Science'. Science 324(5923), pp. 85-89. DOI: 10.1126/science.1165620. https://doi.org/10.1126/science.1165620
  3. Miller, Boaz; Record, Isaac (2013). 'Justified Belief in a Digital Age: On the Epistemic Implications of Secret Internet Technologies'. Episteme 10(2), pp. 117-134. DOI: 10.1017/epi.2013.11. https://doi.org/10.1017/epi.2013.11
  4. Winsberg, Eric (2010). 'Science in the Age of Computer Simulation'. University of Chicago Press. DOI: 10.7208/chicago/9780226902050.001.0001. https://doi.org/10.7208/chicago/9780226902050.001.0001
  5. Freiman, Ori; Miller, Boaz (2019). 'Can Artificial Entities Assert?'. In: Sanford Goldberg (ed.), The Oxford Handbook of Assertion. Oxford University Press, pp. 413-434 (chapter first published online 2019; printed volume dated 2020). DOI: 10.1093/oxfordhb/9780190675233.013.36. https://doi.org/10.1093/oxfordhb/9780190675233.013.36
  6. Freiman, Ori (2022). 'Making sense of the conceptual nonsense 'trustworthy AI''. AI and Ethics 3(4), pp. 1351-1360 (article first published online 2022; issue dated 2023). DOI: 10.1007/s43681-022-00241-w. https://doi.org/10.1007/s43681-022-00241-w
  7. Grindrod, Jumbly (2019). 'Computational beliefs'. Inquiry 68(10), pp. 3169-3190 (article first published online 2019; issue dated 2025). DOI: 10.1080/0020174X.2019.1688178. https://doi.org/10.1080/0020174X.2019.1688178
  8. Durán, Juan M.; Formanek, Nico (2018). 'Grounds for Trust: Essential Epistemic Opacity and Computational Reliabilism'. Minds and Machines 28(4), pp. 645-666. DOI: 10.1007/s11023-018-9481-6. https://doi.org/10.1007/s11023-018-9481-6
  9. King, Ross D.; Schuler Costa, Vlad; Mellingwood, Chris; Soldatova, Larisa N. (2018). 'Automating Sciences: Philosophical and Social Dimensions'. IEEE Technology and Society Magazine 37(1), pp. 40-46. DOI: 10.1109/MTS.2018.2795097. https://doi.org/10.1109/MTS.2018.2795097
  10. Königs, Peter (2022). 'Artificial intelligence and responsibility gaps: what is the problem?'. Ethics and Information Technology 24(3), article 36. DOI: 10.1007/s10676-022-09643-0. https://doi.org/10.1007/s10676-022-09643-0
  11. Freiman, Ori (2023). 'Analysis of Beliefs Acquired from a Conversational AI: Instruments-based Beliefs, Testimony-based Beliefs, and Technology-based Beliefs'. Episteme 21(3), pp. 1031-1047 (article first published online 2023; issue dated 2024). DOI: 10.1017/epi.2023.12. https://doi.org/10.1017/epi.2023.12
  12. Freiman, Ori (2024). 'AI-Testimony, Conversational AIs and Our Anthropocentric Theory of Testimony'. Social Epistemology 38(4), pp. 476-490. DOI: 10.1080/02691728.2024.2316622. https://doi.org/10.1080/02691728.2024.2316622
  13. Fricker, Elizabeth (2025). 'On the metaphysical and epistemic contrasts between human and AI testimony'. Inquiry 69(6), pp. 2859-2884 (article first published online 2025; issue dated 2026). DOI: 10.1080/0020174X.2025.2553297. https://doi.org/10.1080/0020174X.2025.2553297
  14. Butlin, Patrick; Viebahn, Emanuel (2025). 'AI Assertion'. Ergo 12. DOI: 10.3998/ergo.7960. https://doi.org/10.3998/ergo.7960
  15. Vallor, Shannon; Vierkant, Tillmann (2024). 'Find the Gap: AI, Responsible Agency and Vulnerability'. Minds and Machines 34(3), article 20. DOI: 10.1007/s11023-024-09674-0. https://doi.org/10.1007/s11023-024-09674-0
  16. Humphreys, Paul (2019). 'Templates, Opacity, and Simulations'. In: Philosophical Papers (single-author collected volume of Humphreys' papers). Oxford University Press, pp. 61-80. DOI: 10.1093/oso/9780199334872.003.0006. https://doi.org/10.1093/oso/9780199334872.003.0006
  17. Carter, J. Adam; Clark, Andy; Kallestrup, Jesper; Palermos, S. Orestis; Pritchard, Duncan (eds.) (2018). 'Extended Epistemology'. Oxford University Press. DOI: 10.1093/oso/9780198769811.001.0001. https://doi.org/10.1093/oso/9780198769811.001.0001
  18. Krenn, Mario; Champion, Heather (2026). 'Philosophy of Autonomous Science: Ten Questions for the Coming Age of Artificial Scientists'. Daedalus 155(1-2), pp. 335-349. DOI: 10.1162/daed.a.991. https://doi.org/10.1162/daed.a.991
  19. Ratti, Emanuele (2026). 'Epistemic Control and the Normativity of Machine Learning-Based Science'. arXiv:2601.11202. https://arxiv.org/abs/2601.11202