> ## Content Index > Fetch the complete content index at: https://qadmus.ai/llms.txt > Use this file to discover other available public pages before exploring further. # How Do We Know the Machine Is Right? - URL: https://qadmus.ai/how-do-we-know-the-machine-is-right/ - Published: 2026-10-13T14:40:54.000Z - Updated: 2026-10-14T12:33:40.000Z - Description: AI in healthcare is moving fast, but how do we actually know it's right? Explore why academic health systems must look beyond ethics and ask a deeper question: What counts as "knowing"? Introducing a series on epistemology and responsible clinical AI. - Author: Cadmus B Sorrell - Tags: AI Governance, Philosophy, Epistemology, Ethics, AI Ethics, Philosophy of AI, Knowledge and Actions in Healthcare AI Why academic health systems must lead on AI safety and epistemology Oct 13, 2026 - Cadmus Sorrell ## A score, a nurse, and a daughter > Clinical Scene > It is 3 a.m. on a medical floor. A sepsis model flags a 74-year-old woman as high risk. The alert lands on the nurse's screen while the patient sleeps, her vital signs close to her baseline. > Three people now hold three kinds of knowledge about her. The model holds quantitative risk, compiled from structured data in the electronic health record. The nurse holds tacit knowledge, refined over twelve years of clinical observation, and her sense is that this patient looks fine. The patient's daughter, asleep in the chair, holds qualitative history — unrecorded data that never reached the chart. She knows her mother was confused at dinner in a way she has never been before, and no one wrote it down. Who knows best? The question is practical. Beneath it is an older one: What counts as knowing something at all? Philosophers have argued over this for more than two millennia. The scene is a composite, but each part of it is real. When researchers tested one widely used sepsis model at a health system that did not build it, the model missed about two thirds of the patients who developed sepsis, while alerting on nearly one in five of all patients admitted (Wong et al., 2021). Nurses have long described sensing that a patient is turning for the worse before the numbers move. Families notice changes that never reach the chart. Each kind of knowing has a claim, and each can be wrong. This series is about that question. Before we ask what AI in health care ought to do, we need to ask how we know what it is doing, whether it works, and whom it fails. ## Why start with knowledge rather than ethics Most writing on responsible AI starts with ethics. It asks whether a model is fair, whether it respects patient autonomy, and whether its benefits outweigh its harms. Those are the right questions. Readers in academic health systems know them from Beauchamp and Childress, whose four principles of autonomy, beneficence, nonmaleficence, and justice have shaped clinical ethics for more than four decades (Beauchamp & Childress, 2019). Yet each of those principles leans on a claim about knowledge. Judging a model fair requires knowing how it performs across patient groups. Protecting autonomy through informed consent requires knowing the risks well enough to explain them. Weighing benefit against harm requires knowing both. When the knowledge is weak, the ethics built on it is weak too, however sound the principles. Clinicians already work this way. No one writes a treatment plan before making a diagnosis. > Ethics without epistemology is a treatment plan with no diagnosis behind it. It may be well meant, and it may even help, but no one can say why. The research literature makes the same point. In a widely cited review, Mittelstadt and colleagues sorted the concerns raised by algorithms into six types: inconclusive evidence, inscrutable evidence, and misguided evidence, alongside unfair outcomes, transformative effects, and traceability (Mittelstadt et al., 2016). The first three are epistemological. The last three grow, in part, from the first three. So this series takes the slower road. It spends its first long stretch on epistemology alone, then turns to ethics. By the time we reach deontology, utilitarianism, virtue ethics, and care ethics, we should have firm ground to stand on. ## What epistemology asks Epistemology is the branch of philosophy that studies knowledge. It asks what knowledge is, how we get it, what makes a belief justified, and who counts as a knower. The classic starting point comes from Plato's *Theaetetus*, which tests the idea that knowledge is true belief backed by a sound account. In rough form, that view held for centuries as "justified true belief." Then in 1963 Edmund Gettier published a three-page paper with cases where a person holds a justified, true belief and still seems not to know (Gettier, 1963). The person was right, but by luck. Medical AI has its own Gettier cases. Researchers who trained a deep learning model to find pneumonia on chest X-rays found it could also tell which hospital an image came from (Zech et al., 2018). Because pneumonia rates differed between hospitals, the model could lean on where an image was taken instead of what the lungs showed. On some test sets, it satisfied the criteria for justified true belief, yet its "knowledge" was of hospital site tags, not pneumonia. That gap between a right answer and real knowledge is the subject of this series. Each school of epistemology draws the line in a different place. Some say a reliable method is enough. Some demand reasons a person can inspect. Some say knowledge belongs to communities, not to individuals. Some say the knower's place in the world shapes what can be known. Each view leads to different rules for testing, approving, and watching AI in the clinic. ## Why academic health systems Academic health systems are a fitting place to test these ideas, because knowledge is their whole business. They exist to care for patients, teach the next generation of clinicians, and create new medical knowledge. Each of those missions now runs partly through AI. A model that drafts a clinical note shapes what the record says happened. A model that predicts readmission shapes who gets a follow-up call. A model used to build a research cohort shapes what the next study finds. When these systems err, the error does not stay in one chart. It travels into teaching cases, quality reports, and published papers. Academic health systems do not merely use knowledge; they produce it. When an AI system enters here, it enters the pipeline through which medical knowledge flows to the wider world. They also have the people to take these questions seriously. Ethics committees, institutional review boards, AI governance committees, and informatics teams already decide what counts as enough evidence. Mostly, they do so without naming the theory of knowledge they are using. This series aims to name it, so those choices can be argued in the open. ## How the series is built The series has five parts, followed by a bridge into ethics: - **Part I: The Practitioner's Home Ground.** The frameworks clinical and AI teams already inhabit: empiricism, Bayesian updating, reliabilism, rationalism, pragmatism, testimony, and model opacity. - **Part II: Knowledge as Social and Relational.** Moving beyond individual minds to address social epistemology, epistemic injustice (Fricker, 2007), feminist standpoint theory, Ubuntu, Indigenous epistemologies, and trust. - **Part III: Justification Under Strain.** How classic philosophical tensions (foundationalism vs. coherentism, internalism vs. externalism) drive live clinical AI debates like explainability. - **Part IV: Traditions Beyond the West.** Non-European traditions (Nyāya, Buddhism, Jainism, Confucianism, Daoism, Islamic rationalism, African/Global South thought, Zen, and the Kyoto School) that challenge Western analytical assumptions. - **Part V: Prominent but Poor Fits.** Epistemic modes that shape millions of lives but fit AI governance poorly: radical skepticism, intuition, relativism, and authority/revelation. A bridge post then names the theories of knowledge hidden inside utilitarian and Kantian ethics, and hands off to the ethics series. Each post follows the same seven-part shape. It opens with a clinical scene, states the school's core claim in plain terms, shows how the school appears in AI practice, weighs its strengths and weaknesses, and puts it through a health care stress test. It closes by handing off to the next post. Each also carries a rough rating of how well the school fits responsible AI in health care: **strong**, **mixed**, or **weak**. The ratings are a guide, not a verdict, and readers should argue with them. ## What this series will not do It will not crown a winner. No single theory of knowledge is enough for health AI, and anyone who claims one is selling something. The useful question is which school helps with which decision, and where each stops helping. It will not be a literature review. Each post makes an argument, cites the work it draws on, and invites disagreement. Every source that carries weight is listed, so readers can check it. And it will not treat philosophy as decoration. The aim is practical. A governance committee that knows it reasons like a reliabilist can ask what reliabilism misses. A chief medical information officer who sees that a vendor's claims rest on testimony can ask who stands behind that testimony. Naming the theory brings the blind spot into view. ## How to use these posts The series is written for people who govern, build, and judge AI in academic health systems. That includes ethicists, informatics leaders, quality and safety staff, chaplains and other members of ethics committees, clinicians on governance committees, researchers, and the executives who fund all of them. Each post ends with one question to bring to your own AI governance work. The questions are meant for a meeting agenda, not a seminar. Take one to the committee that approves your next model, and listen for which theory of knowledge the room is using. > **Governance Question** > When your committee approved its last AI model, what did it accept as proof that the model works, and who had the authority to decide that proof was sufficient? ## Next, empiricism The series begins where most of health AI already lives. Empiricism holds that knowledge comes from experience and observation. It gave medicine the clinical trial, and it gives AI the validation study and the audit. It is a strong school, and the next post gives it full credit. It carries a problem David Hume named in 1748: no number of past observations proves the next case will follow the same pattern (Hume, 1748/1999). In AI, that old problem has a new name: distribution shift. It explains how the sepsis model in our opening scene could perform well in the hospitals that built it and poorly in the one that bought it. Post 2 begins there. ## References 1. Beauchamp, T. L., & Childress, J. F. (2019). *Principles of Biomedical Ethics* (8th ed.). Oxford University Press. 2. Fricker, M. (2007). *Epistemic Injustice: Power and the Ethics of Knowing*. Oxford University Press. 3. Gettier, E. L. (1963). Is justified true belief knowledge? *Analysis, 23*(6), 121–123. 4. Hume, D. (1999). *An Enquiry Concerning Human Understanding* (T. L. Beauchamp, Ed.). Oxford University Press. (Original work published 1748) 5. Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms: Mapping the debate. *Big Data & Society, 3*(2). 6. Plato. (1990). *Theaetetus* (B. Williams, Ed.; M. J. Levett, Trans.). Hackett. 7. Wong, A., Otles, E., Donnelly, J. P., et al. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. *JAMA Internal Medicine, 181*(8), 1065–1070. 8. Zech, J. R., Badgeley, M. A., Liu, M., et al. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs. *PLOS Medicine, 15*(11), e1002683.