The DOI matches its Crossref record; 83 of the 83 references it cites were found in a registry or an index; 8 of the 46 sources its claims cite were checked for support.
The Epistemic Cost of Opacity: How the Use of Artificial Intelligence Undermines the Knowledge of Medical Doctors in High-Stakes Contexts
Open accessAbstract
Abstract Artificial intelligent (AI) systems used in medicine are often very reliable and accurate, but at the price of their being increasingly opaque. This raises the question whether a system’s opacity undermines the ability of medical doctors to acquire knowledge on the basis of its outputs. We investigate this question by focusing on a case in which a patient’s risk of recurring breast cancer is predicted by an opaque AI system. We argue that, given the system’s opacity, as well as the possibility of malfunctioning AI systems, practitioners’ inability to check the correctness of their outputs, and the high stakes of such cases, the knowledge of medical practitioners is indeed undermined. They are lucky to form true beliefs based on the AI systems’ outputs, and knowledge is incompatible with luck. We supplement this claim with a specific version of the safety condition on knowledge, Safety*. We argue that, relative to the perspective of the medical doctor in our example case, his relevant beliefs could easily be false, and this despite his evidence that the AI system functions reliably. Assuming that Safety* is necessary for knowledge, the practitioner therefore doesn’t know. We address three objections to our proposal before turning to practical suggestions for improving the epistemic situation of medical doctors.
Overview
CoChat Summary
TL;DR
The paper argues that medical doctors who rely on opaque AI outputs in high-stakes contexts may form true beliefs but still lack knowledge because their beliefs fail a perspective-relative safety condition, Safety*. Using a hypothetical breast-cancer recurrence case, it claims opacity, correctness-blindness, potential malfunction, and high stakes make the doctor’s true belief lucky and therefore not knowledge.
Key findings
- In the “Cancer Risk” case, an opaque AI system predicts that patient P is not at risk—below a threshold of 0.05—and MD forms a true belief on that basis. The authors argue MD does not know because, relative to his perspective, the belief could easily be false if the system were malfunctioning; it is therefore not Safe*.
- Four features undermine MD’s knowledge: the system’s opacity, MD’s inability to check correctness directly, the real possibility of AI flaws/malfunctions/biases, and the high stakes for P and MD. MD’s evidence that the system works reliably does not remove the doubt that defeats knowledge.
- The paper blocks three objections: overgeneralization to perceptual expertise, expert testimony, CT scanners, and RCTs; undergeneralization from the breast-cancer case; and the claim that knowledge is practically irrelevant. It argues knowledge matters for moral/legal responsibility and informed consent.
- Practical remedies suggested are independent sources and “broken-leg” override, explainable AI or expert understanding of the system, and reliable validation/certification structures.
Methods
- Conceptual and philosophical analysis, not an empirical study. The paper uses a hypothetical medical case, “Cancer Risk,” plus the “Broken Clock” epistemological analogy.
- It develops and applies Safety*: “If S were to believe that p, then relative to her perspective, it could not easily be false that p,” drawing on Whiting.
- It addresses three objections and then proposes practical measures; the paper states there are no data, materials, or code to be made available.
Limitations
- The paper states no formal limitations section; its Data Availability statement says there are no data, materials, or code to be made available.
- Evident from the text: the argument rests on a hypothetical case and philosophical intuitions rather than empirical measurement, and its conclusion depends on accepting Safety* as necessary for knowledge and on pragmatic encroachment about high stakes.
- Evident from the text: the scope is limited to high-stakes, correctness-blind medical contexts; the authors acknowledge that complete understanding of AI inner workings remains improbable and that generalization depends on shared structural features.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 5 Not valid · 19 Not checked · 58 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Importantly, it is often not only the (global) inner workings of such systems that are hard to understand, but also the (local) causes of their individual outputs, for instance, why a certain CT scan was classified as depicting a tumor, or why a certain medical treatment was recommended for a patient”
What the sources say
- Supportedcites Reyes et al., 2020 source
The Introduction’s “Global versus Local” section distinguishes explanations of patterns driving a model’s predictions from explanations of why it made a specific prediction for a given input. The paper’s Essentials also says radiology AI systems can be difficult for humans to interpret, supporting the claim that both levels can be hard to understand. (LLM, conf=0.91)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“We can say that some AI systems are, to a certain extent, opaque. Sometimes the opacity of a system is due merely to the fact that the system is proprietary, but other times the reason is that its model involves sub-symbolic representations incomprehensible even to experts, or that it is too large or too complex to grasp”
What the sources say
- Not checked — no full text of the sourcecites Bathaee, 2018
- Partly supportedcites Ghassemi et al., 2021 sourceExperimental — accuracy not yet measured
The Introduction says that many modern AI models are too complex and high-dimensional to be easily understood, supporting opacity due to complexity. It does not identify proprietary status or sub-symbolic representations as reasons for opacity. (LLM, conf=0.88)
- Partly supportedcites Mann et al., 2023 sourceExperimental — accuracy not yet measured
Our copy supports most of the sentence: Foreign Representation (Architectural Opacity, p. 2) attributes opacity to AI systems using distributed, subsymbolic representations generated in training; Complexity (same section, p. 2) covers systems 'complex because of scale', 'especially severe for the millions of parameters of some neural networks'; and Intentional Concealment (Socio-Technical Opacity, p. 4) covers private developers seeking opacity to protect intellectual property or secure economic advantage, with 'relying on open-source software instead' as a remedy — the closest the paper comes to 'proprietary'. However, the added qualifier that these sub-symbolic representations are 'incomprehensible even to experts' is not the paper's position: it says representations are foreign 'to humans' while listing XAI strategies (saliency maps, network dissection, concept activation vectors, p. 3) to map them onto human concepts, and Section II (p. 1) explicitly holds opacity to be stakeholder-dependent ('a system can be transparent for an expert but opaque for a layperson'). The paper also never uses the term 'proprietary', treating that motive as one instance of a broader intentional-concealment category. (LLM, conf=0.78)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Assuming that Safety* is necessary for knowledge, the practitioner therefore doesn't know.”
plainSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We argue that, relative to the perspective of the medical doctor in our example case, his relevant beliefs could easily be false, and this despite his evidence that the AI system func -tions reliably.”
hedgedSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“So, given that it is a matter of luck that MD's belief is true, he doesn't know that P is not an at-risk patient.”
- Also stated in: § 2 The Intuitive Case for Undermined Knowledge · p. 7 · plain
“So his belief is true only by luck, and he does not have knowledge.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Also stated in: § 2 The Intuitive Case for Undermined Knowledge · p. 7 · plain
37 more claims
Create a free account to see all 89 claims and to verify the 2 references still open.
- Probably not a substantial claim. 34
- It may state two separate claims as one. 19
Citations
Cites
High-performance medicine: the convergence of human and artificial intelligence
Topol, E. J. (2019) · Nature Medicine · doi:10.1038/s41591-018-0300-7 · Cited by 10,450 · cited 3×
“Also see Topol ( 2019 ).)”
Dissecting racial bias in an algorithm used to manage the health of populations
Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019) · Science · doi:10.1126/science.aax2342 · Cited by 7,137 · cited 1×
“( For an illustra - tive case, see Obermeyer et al., 2019 .)”
(2000) · Cited by 5,055
Is Justified True Belief Knowledge?
Gettier, E. L. (1963) · Analysis · doi:10.1093/analys/23.6.121 · Cited by 3,127 · cited 1×
“She argues that belief with such contents constitutes knowledge under the same conditions as belief with non-probabilistic contents, e.g. when the subject is not in a Gettier case or a fake barn case (Gettier, 1963 ; Goldman, 1976).”
Williamson, T. (2002) · doi:10.1093/019925656x.001.0001 · Cited by 2,788
Discrimination and Perceptual Knowledge
Goldman, A. I. (1976) · The Journal of Philosophy · doi:10.2307/2025679 · Cited by 1,870 · cited 1×
“She argues that belief with such contents constitutes knowledge under the same conditions as belief with non-probabilistic contents, e.g. when the subject is not in a Gettier case or a fake barn case (Gettier, 1963 ; Goldman, 1976).”
The false hope of current approaches to explainable artificial intelligence in health care
Ghassemi, M., Oakden-Rayner, L., & Beam, A. L. (2021) · The Lancet Digital Health · doi:10.1016/s2589-7500(21)00208-9 · Cited by 1,724 · cited 3×
“…of a system is due merely to the fact that the system is proprietary, but other times the reason is that its model involves sub-symbolic representations incomprehensible even to experts, or that it is too large or too complex to grasp (Bathaee, 2018 ; Ghassemi et al., 2021 ; Mann et al., 2023 ).”
Human-computer collaboration for skin cancer recognition
Tschandl, P., Rinner, C., Apalla, Z., Argenziano, G., Codella, N., Halpern, A., Janda, M., Lallas, A., Longo, C., Malvehy, J., et al. (2020) · Nature Medicine · doi:10.1038/s41591-020-0942-0 · Cited by 1,024 · cited 1×
“As examples, see the studies cited in Bjerring and Busch ( 2021 , 350) as well as Tschandl et al. ( 2020 ) or Tu et al. ( 2024 ).”
How to Defeat Opposition to Moore
Sosa, E. (1999) · Noûs · doi:10.1111/0029-4624.33.s13.7 · Cited by 969
Pritchard, D. (2005) · doi:10.1093/019928038x.001.0001 · Cited by 874
Cited by
“Does Black Box AI In Medicine Compromise Informed Consent?”
Director, S. (2025) · Philosophy & Technology · doi:10.1007/s13347-025-00860-1 · Cited by 25
Herrera, F., & Calderón, R. (2026) · Technology in Society · doi:10.1016/j.techsoc.2026.103302 · Cited by 36
The end of theory? AI and ignorance in financial markets
Svetlova, E., & Arnoldi, J. (2026) · Finance and Society · doi:10.1017/fas.2026.10032 · Cited by 2
“…in finance), and indeed also a literature on ignorance and AI, most of which is highlighting the opacity of AI systems (Emmert-Streib, Yli-Harja, and Dehmer, 2020; Boge, 2022; Borch and Min, 2022; Khan and Vice, 2022; Duede, 2023; Carloni, Berti, and Colantonio, 2025; Schmidt et al., 2025).”
Med-ViX-Ray: Enhancing explainable chest X-ray analysis with clinical knowledge graphs
Cieri, M., & Palomba, F. (2026) · Computer Methods and Programs in Biomedicine · doi:10.1016/j.cmpb.2026.109313 · Cited by 2
Approaches to AI Integration in Education and Healthcare
Evangeline, S. I. (2025) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-5072-1.ch011 · Cited by 1
Metadata
- Authors
- Schmidt, E., Putora, P. M., & Fijten, R.
- Affiliations
- TU Dortmund University (DE)
- University of Bern (CH)
- Maastricht University Medical Centre (NL)
- Corresponding author
- Schmidt, Eva
- Venue
- Philosophy & Technology
- Published
- 2025-01-13
- Volume
- 38
- Issue
- 1
- ISSN
- 2210-5433 (ISSN-L)
- 2210-5441
- OpenAlex
- W4406334317
- Semantic Scholar
- 839345de6a83ea740bbb45af20641cf8916d0f6e
- Cited by
- 14
- References
- 77
- Citations by year
- 2026: 11 · 2025: 3
- FWCI
- 111.65
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Philosophy & Technology (publishedVersion, cc-by)
- Research Publications (Maastricht University) (publishedVersion, cc-by)
- Funders
- Technische Universität Dortmund
- Topics
- Social Sciences › Arts and Humanities › Philosophy › Epistemology, Ethics, and Metaphysics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Social Sciences › Decision Sciences › Information Systems and Management › Knowledge Management and Technology
- Keywords
- artificial intelligence in healthcare, algorithmic opacity, epistemic luck, breast cancer recurrence prediction, high-stakes decision making, trust in AI outputs, epistemic justification
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The Epistemic Cost of Opacity: How the Use of Artificial Intelligence Undermines the Knowledge of Medical Doctors in High-Stakes Contexts. https://app.cochat.ai/papers/10.1007/s13347-024-00834-9
- The paper
- Schmidt, E., Putora, P. M., & Fijten, R. (2025). The Epistemic Cost of Opacity: How the Use of Artificial Intelligence Undermines the Knowledge of Medical Doctors in High-Stakes Contexts. Philosophy & Technology. https://doi.org/10.1007/s13347-024-00834-9
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.