The Doctor and the Machine

A patient arrives with a problem that does not announce its name. The symptoms point in several directions. The laboratory results narrow some possibilities and open others. An imaging study adds another clue. A physician has limited time, incomplete information, and a decision that may matter immediately. Now add an AI system that can read the chart, compare patterns across thousands of cases, suggest a list of possible diagnoses, called a differential diagnosis, identify a drug interaction, and recommend the next test.

The machine may be useful. It may even be right when the physician is wrong. That is where the easy argument ends.

Medicine forces the delegation question into its sharpest form because accuracy, authority, and accountability are not the same thing. A system can produce a better answer without becoming the party that is allowed to decide. A physician can remain legally or professionally responsible while having too little information, time, or practical power to review the machine meaningfully. A hospital can buy the system, embed it in the workflow, and then behave as if every consequence belongs to the clinician who happens to click the final button.

AI has already entered medicine, and its capabilities are improving. The governing question is what role the machine occupies inside the clinical chain, what decisions may be delegated, and where human judgment must remain real rather than ceremonial.

The Clinical Chain

A diagnosis looks like a single conclusion only after the work has been compressed. Before that conclusion sits a chain of choices: which information to collect, which symptoms deserve weight, which possibilities belong in the differential, which tests to order, how to interpret the results, when to escalate, what treatment options to consider, how to explain uncertainty, and who finally decides what happens next.

AI can help at almost every step. It can summarize a record, flag an unusual image, suggest possible diagnoses or tests, and draft a treatment plan. It can also warn that a treatment may be unsuitable for a particular patient or flag signs that a patient’s condition is worsening. These uses need different rules. Drafting words for review gives the machine a different role from letting its output trigger care.

That distinction matters because the machine can be right without being responsible. Responsibility belongs to the people and institutions that choose the system, define its role, decide when its output can trigger action, determine who must review it, and live with the consequences. Principle 1 applies with unusual force in medicine: accountability does not transfer merely because a machine produced the recommendation.

The clinical chain also prevents the opposite mistake. Keeping a human somewhere in the workflow does not prove that the delegation is responsible. If the physician sees only a score, has thirty seconds to react, does not know which data drove the recommendation, and faces institutional pressure to accept it, the human may be present without being able to exercise independent judgment. That is not meaningful review. It is a signature attached to a process controlled elsewhere.

Capability Is Not Permission

Medicine will increasingly confront an uncomfortable possibility: for some bounded tasks, the machine may outperform the average clinician. That result leaves the delegation question open. Principle 2 separates capability from permission. Competence is evidence about what work a system can perform. Permission is a decision about what authority society, a profession, or an institution is prepared to give it.

The distinction is easier to see outside medicine. A calculator can outperform a person at arithmetic without acquiring authority to approve a loan. A navigation system can find a faster route without acquiring authority to decide where the passenger must go. In medicine the stakes are higher, but the structure is the same. Diagnostic capability and diagnostic authority are different things.

This argument concerns legitimate authority rather than professional territory. A rule requiring a human to make every clinical judgment simply because humans have always made it would become difficult to defend if a machine repeatedly proved safer and more accurate. The Doctrine rejects any presumption that people are always superior decision makers. It asks whether the delegation has a legitimate owner, a defined boundary, real review where review is required, and consequences that can reach the people and institutions controlling the risk.

This means better AI should change some boundaries. A system that has been rigorously validated for a narrow screening task may deserve more autonomous operation than a general model generating an open-ended treatment plan from a complex record. The evidence, the decision, and the consequences should set the boundary, rather than fear of automation or admiration for it.

The Human in the Loop Can Be a Fiction

Healthcare organizations often answer concerns about AI with a reassuring phrase: a human remains in the loop. The phrase sounds like accountability and can describe something real. It can also describe almost nothing.

A reviewer needs four things: authority over the case; enough evidence and context to understand the recommendation; enough time and competence to disagree; and practical power to override the output without punishment. If any of these is missing, the person may be confirming the machine’s answer without being able to judge it independently.

The FDA’s January 2026 guidance on clinical decision support makes a closely related distinction. For certain software functions to fall outside the medical-device definition under the statutory clinical decision support criteria, the healthcare professional must be able to independently review the basis for the recommendation rather than rely primarily on the software. Compared with the 2022 guidance, the January 2026 revision moved toward less regulation at the margin: it added enforcement discretion for certain functions that provide a single clinically appropriate recommendation and placed time-criticality inside the independent-review analysis. The core boundary remained whether the professional can independently review the basis for the recommendation. FDA points to intended use, required inputs, algorithm development and validation, and patient-specific knowns and unknowns as information that can make that review possible.

The final guidance also retains an explicit warning about automation bias, the tendency to over-rely on automated suggestions. It describes both errors of commission, following incorrect advice, and errors of omission, failing to act because the system did not prompt action. FDA also connects automation bias to urgent decisions, where limited time can make independent review harder. A system can therefore be technically supervised while functionally driving the decision.

Principle 4 says review must be real. Medicine shows why. Oversight is not the number of human clicks between a model and an outcome. It is the presence of an informed person with authority, context, time, and a genuine ability to say no.

The Confidence Problem

Clinical decisions are made under uncertainty, and uncertainty is easy to hide once an answer arrives in polished language. A diagnosis is rarely just a label. It carries a level of confidence, a set of alternatives, missing information, and a judgment about whether the risk of waiting is greater than the risk of acting. An AI system that produces the right-looking sentence can compress all of that ambiguity into something that feels more settled than the evidence actually is.

Humans have the same weakness. Physicians can anchor on an early diagnosis, overlook contradictory evidence, or become too confident in a familiar pattern. Overconfidence predates AI. AI changes its scale and presentation. A recommendation may arrive instantly, repeatedly, and with the visual authority of a system embedded in the medical record. That makes the design of uncertainty part of the accountability architecture.

A reviewer should therefore see more than the machine’s preferred answer when the stakes justify it. Useful oversight may require the relevant inputs, important missing data, competing possibilities, limitations of the model, and a clear trigger for escalation. FDA’s current clinical decision support guidance points in this direction by emphasizing patient-specific knowns and unknowns and enough information about the basis of a recommendation for independent review.

This is also where abstention becomes a form of competence. Abstention means declining to give an answer or make a decision when doing so would go beyond the evidence or the system’s role. The information may be incomplete, the data unreliable, or the patient outside the group for which the system was tested. The case may need a human specialist. A boundary is useful partly because it gives the machine somewhere to stop.

Autonomy Can Be Legitimate

Human review can be fake, and autonomy can be misunderstood. A medical system can act autonomously inside a legitimate accountability structure when its authority is narrow, validated, monitored, and connected to an escalation path. The question is what the system is authorized to decide and what happens after it decides.

Diabetic retinopathy screening offers a useful example because the boundary is narrow. In 2018 the FDA granted De Novo classification to IDx-DR, now known as LumineticsCore, a prescription device designed to automatically detect more-than-mild diabetic retinopathy in specified adults with diabetes using designated retinal imaging. Its authorization covered a narrow screening task rather than every eye disease, treatment choice, or the rest of ophthalmic care. Its autonomy sits inside a written clinical boundary.

Johns Hopkins Medicine later deployed the system in community primary-care sites. A 2026 study of 3,745 adults with diabetes compared patients referred to specialist eye care through ordinary clinician referral with patients referred after a positive or non-diagnostic autonomous AI screening result. The researchers reported an association between the AI pathway and increased presentation to specialist care among African-American patients. They also emphasized that the analysis was exploratory and could not establish causation.

That case matters because the AI did not merely advise a physician about what an image might show. Within its authorized task, the system produced a diagnostic screening result that could trigger referral. Yet the surrounding accountability architecture remained visible: a defined patient population, a defined disease target, specified imaging, a regulated device, a referral pathway, and downstream specialist care.

This is a better way to think about autonomy than asking whether a doctor touched every output. A bounded machine can sometimes act more responsibly than a nominally supervised system whose boundaries are vague. Principle 3 requires a boundary, not ritual human interference with every competent act.

The Reserved Line

Medicine also contains decisions for which technical performance is only part of the legitimacy question. The Doctrine’s Reserved category, developed in Part Six, covers decisions whose legitimacy depends on a human making them, especially where a decision directly and finally determines a fundamental interest such as life, liberty, livelihood, health, reputation, or a child’s development. One explicit example is disclosure of a terminal diagnosis.

Capability is separate from legitimacy here. A future system may identify a terminal condition earlier and estimate prognosis better than many clinicians. The Reserved question begins after capability. Who is authorized to weigh uncertainty, understand the patient’s circumstances, decide how the information should be communicated, answer questions, respond to emotion, and remain answerable for what follows?

Reserved is narrower than medicine as a whole. The tiered test this book develops in Part Six reserves a human signature for irreversible, high-stakes decisions in Tier 3 and a still narrower Reserved category for decisions whose legitimacy depends on a human making them. A bounded screening classification may be delegated. A medication-interaction alert may be automated. A routine recommendation may sit in a supervised category that requires human review. The important distinction is between high stakes, which can require a human signature, and Reserved legitimacy, which requires the human act of judgment itself.

Reserved should remain smaller and harder. It applies when the human act itself contributes to legitimacy, rather than merely when the subject matter is important. Final decisions to withdraw life-sustaining treatment, communicate a terminal diagnosis, or make other choices in which values and irreversible consequences are inseparable from the decision are the clearest medical candidates. AI may inform those decisions deeply. Final accountable authority remains human because legitimacy requires it.

Principle 7 therefore requires more than a name in the chart. For a Reserved decision, independent human judgment must come from an authorized and competent decision owner with the evidence, context, and practical authority to disagree. The human signature is not a ceremony performed after the machine has already settled the matter.

A current regulatory debate makes the line concrete. In 2026 the Washington Medical Commission proposed an interpretive statement reading existing state law to mean that only natural persons can hold a license to practice medicine. A companion policy proposal stated that AI may inform care but may not independently and autonomously assess, diagnose, treat, or prescribe. The American Medical Association supported that licensure reading while asking the Commission to revise separate language that would place full responsibility on physicians even when an AI tool was embedded or mandated by an employer, health system, or payer. At this stage these Washington materials are proposals, not final statewide rules. Even so, they show a medical regulator drawing the human-signature boundary in the language of licensure.

Who Owns the Error?

A patient harmed by an AI-influenced decision will rarely be harmed by one actor alone. The developer may have trained the model on unrepresentative data. The hospital may have deployed it in a population for which it was not validated. The electronic record may have supplied stale information. Administrators may have made use mandatory. The physician may have ignored contradictory evidence or accepted the output too quickly. Monitoring may have failed to detect performance drift. A useful accountability system has to follow control rather than simply attach the entire consequence to the last person in the chain.

The developer is responsible for how the system is built and tested, what its limits are, and how honestly those limits and later changes are explained. The health system is responsible for choosing the tool, checking it with the patients and conditions where it will be used, connecting it to daily care, and giving staff the training, time and authority to use it safely. That includes clear rules for when to seek help and when to stop using it.

The clinician is responsible for judgment about the individual patient, within the choices and information the clinician actually has. The patient brings personal values and makes informed choices. Patients should not be made answerable for hidden defects in the software or decisions they could not see or control.

That division is an accountability map rather than a legal liability formula. Courts, statutes, contracts, regulation, and professional rules will draw legal lines differently across jurisdictions and products. Responsibility should track the parts of the system each actor can actually control.

The American Medical Association has argued for a similar control-based approach in its policy advocacy. In 2026 it objected to proposals that would place full responsibility on a physician even when an AI tool was mandated by an employer or failed in ways a reasonable clinician could not detect. The AMA’s position is a stakeholder view, not settled law, but the underlying question is exactly the one Principle 5 raises: architecture shapes accountability.

If a health system requires use of a model, hides its limitations, gives clinicians no meaningful way to inspect or override it, and then treats the physician as the sole accountable party, the organization has engineered responsibility away from control. That is not an accountability system. It is a liability shield.

Current Evidence: The Doctor-Machine Handoff

Reviewed October 6, 2026. UK lifecycle oversight evidence added October 7, 2026. Automated consent-review evidence added October 15, 2026. Medical execution evidence and FDA deadline verified October 20, 2026.

A randomized clinical trial published in JAMA Network Open in 2024 gives the handoff problem a clean experimental form. Fifty physicians in internal medicine, family medicine, and emergency medicine worked through difficult clinical vignettes with either conventional resources or conventional resources plus access to GPT-4 through ChatGPT Plus. Median diagnostic reasoning scores were 76% with the LLM and 74% with conventional resources, a difference that was not statistically significant. In an exploratory comparison, the LLM working alone scored a median 92%, 16 percentage points above the conventional-resource physician group.

That result is more interesting than either “AI beats doctors” or “AI does not help doctors.” In this small study, access to a stronger standalone performer did not automatically improve the human-machine team. The physicians also received little training with the tool, a limitation the paper itself notes. The bottleneck was not merely model capability. It was integration, prompting, trust, workflow, and the ability of physicians to use the tool effectively. Medicine has its own conversion gap.

The capability frontier has continued to move. In May 2026, researchers reported in Nature Medicine that a multimodal version of Google’s AMIE system outperformed primary-care physicians on diagnostic accuracy and on 29 of 32 evaluation axes in 105 simulated telehealth scenarios, comparing AMIE against 19 board-certified primary-care physicians and using 18 specialists as raters. The study was randomized and blinded, but it was an exploratory simulation rather than a real-world clinical trial, and the system was developed and funded by Google. It is evidence of capability, not proof of safe autonomous practice.

Regulation is moving at the same time. The FDA’s final January 2026 clinical decision support guidance emphasizes the conditions under which a professional can independently review a recommendation, while narrowing the range of software the agency treats as a medical device. In August 2026 the agency issued a separate discussion paper seeking public feedback on risk assessment, competency-based premarket evaluation, postmarket monitoring, foundation models, and agentic AI systems. FDA requested comments by October 19, 2026 under docket FDA-2026-N-7874. The August document is a discussion paper, not a new rule. Its significance is that regulators are now confronting systems whose outputs and behavior may be more open-ended than traditional medical software.

On October 6, 2026, the UK government accepted all 44 recommendations of its National Commission into the Regulation of AI in Healthcare. Its response calls for assessment and monitoring throughout an AI medical device’s working life, and the next phase of the MHRA’s AI Airlock will test post-market surveillance and lifecycle regulation. These are commitments and a testing programme, not proof that the proposed reforms already work.

The accountability lesson is that permission needs continuing evidence. A hospital’s approval should identify who watches performance after deployment, what changes trigger another review, and who can suspend use. A tool can change while its original approval remains on file. Oversight must keep pace with the system people are actually using.

The Johns Hopkins diabetic-retinopathy experience points in the other direction from the assumption that every autonomous decision requires a physician to repeat it. The AI system operated within a narrow, regulated screening role and could send positive or non-diagnostic cases into a specialist referral path. The 2026 study found an association with better downstream presentation to eye care among African-American patients, while expressly warning that the observational analysis could not prove causation. The case shows why governance needs tiers rather than a single rule requiring either total automation or total human review.

Utah’s Nolla Health pilot makes graduated delegation concrete. Launched October 5, 2026, it permits AI-issued initial prescriptions for a restricted list of topical acne treatments for Utah adults with mild to moderate acne. The first stage requires two independent physicians to approve every prescription and lasts at least four weeks, covering at least 100 patients. Later stages permit prescribing before review, first with every case reviewed at least weekly, then with monthly review of a minimum 10% sample and full review of adverse events and escalations. Each advance requires written state permission and specified safety results. Cases outside the permitted scope must stop or move to human care. Utah expressly distinguishes this temporary authorization from endorsement of the technology. The lesson is an accountability structure being tested: authority expands only within stated boundaries, with monitoring, escalation, and power to withdraw permission.

On October 3, 2026, the American Academy of Pediatrics recommended that generative AI used in children’s care be evaluated for pediatric populations, with continuing monitoring and clear human oversight. Adult patient data may miss important differences in children’s physiology and development. These are professional recommendations, not proof that a particular system is safe.

The delegation boundary therefore includes who the patient is. A hospital should not assume that evidence from adult care justifies using the same tool with children. Before expanding its role, the institution should establish what was tested, which children were represented, who reviews uncertain results, and who can suspend use when problems appear.

A study published on September 26, 2026, and publicized by Mount Sinai on October 8 tested 20 open-weight language models against unsafe clinical instructions. Across more than 10 million responses to constructed scenarios and cases adapted from hospital records, potentially harmful choices fell from 16.6% without a safety reminder to 10.1% with one. The improvement occurred in 19 of 20 models. These were experimental choices, not observed patient injuries or estimates of risk in deployed care.

Clinical testing therefore needs to ask more than whether an assistant knows the correct answer. It should also test whether the system questions an unsafe request, seeks verification, or escalates to a clinician. A helpful reminder is one control to evaluate. The remaining failures mean it cannot substitute for a validated clinical boundary and competent oversight.

An October 14, 2026 study in npj Digital Surgery tested automated review of 50 simulated surgical-consent conversations against assessments by eight clinicians. One AI judge passed every rubric item, producing high agreement while missing every clinician-identified failure. Combining judges caught more failures, but the authors found no tested arrangement suitable for automatic triage. This was a small simulation, not a trial involving patients.

The safeguard must be tested for the failures it is supposed to catch. An average agreement score can look reassuring when most examples are acceptable. A hospital using AI to review another assistant should measure missed problems and review workload, compare findings from suitably qualified clinicians, and retain human responsibility for the consent process.

In an October 19, 2026 review of medical-AI evaluations, Stanford researchers described agents documenting the correct action without placing the order in the electronic patient record. These controlled tests do not establish how often such failures occur in routine care.

The hospital’s review must therefore check what happened in the record, not only what the assistant wrote. Before expanding an agent’s role, the institution should test whether authorized orders reach the right patient, whether incomplete tasks are flagged, and who acts when work stalls. A convincing note cannot be the evidence that its own plan was carried out.

In a study published in Science on October 22, 2026, Microsoft researchers reported that their diagnostic system reached the correct answer in 83% of complex case-based tests, compared with 36% for physicians using reference resources. Participants gathered information step by step. These controlled results do not establish patient benefit or safe autonomous practice.

Once an assistant helps choose the next test, it shapes the information available for the next decision. A hospital should evaluate that sequence, including unnecessary tests and missed escalation, rather than score only the final diagnosis. Promising performance can justify testing a larger role, while the institution keeps responsibility for patient selection, consent, review and stopping rules.

On October 26, 2026, the Bipartisan Policy Center argued that clinical-AI payment should depend on demonstrated value rather than simply adding billable services. Its policy analysis warns that paying separately for AI that reduces clinician time could increase spending without improving outcomes. It also identifies training, workflow redesign and governance as adoption costs, particularly difficult for under-resourced providers.

The payment arrangement is part of the delegation boundary. Before adopting a clinical tool, a health system should budget for the staff who review uncertain results, investigate failures and decide when use must stop. It should also ask whether payment rewards better care or merely more machine activity. Savings that leave those duties unfunded can make an apparently efficient service difficult to govern.

A randomized trial published in JAMA Network Open on November 4, 2026 assigned 400 pregnant or postpartum participants from racial and ethnic minority groups to a maternal-health chatbot or monthly children’s books. Chatbot access produced no significant improvement in anxiety or depression at 12 months. Symptoms were generally low at enrollment, limiting what the trial could establish about treating clinical illness. The result concerns this intervention and population, not every medical chatbot.

A health service should therefore define the benefit it expects before delegating support to an assistant. More availability or conversation does not establish better care. The institution should measure patient outcomes and preserve access to qualified help when the assistant is insufficient.

Together these findings produce an awkward but useful picture. AI can outperform clinicians in controlled diagnostic tasks. Giving clinicians access to AI does not guarantee that the combined team improves. Narrow autonomous systems can already operate inside real clinical pathways. Regulators are trying to preserve independent review where review is the intended safeguard while also confronting systems that may not fit the old decision-support model. The evidence does not support one universal human-in-the-loop rule.

Accountability Analysis

The clinical accountability chain begins with Principles 1 and 3: the machine does not become the accountable party, and every delegation needs a boundary. Inside that frame, Principle 2 does the first substantive work. A model’s superior score on a benchmark can justify testing or expanding a role, but capability alone cannot authorize every clinical decision that resembles the benchmark. The harder the decision, the more important the boundary between technical competence and legitimate authority.

Principle 4 asks whether review is real. What information must the clinician receive? Is there enough time to evaluate it? Can the clinician understand the basis well enough to disagree? Can the clinician actually override it? A hospital that cannot answer those questions has installed software and hoped professional responsibility will fill the gaps. The four reviewer conditions developed in this chapter are the clinical form of Principle 7’s signer rule: authority over the case, evidence and context, time and competence to disagree, and practical power to override without punishment.

Principle 7 then draws the Reserved line. Some decisions require an independent human decider because the act of human judgment is part of their legitimacy. The developer, institution, and clinician can still use AI extensively around that decision, but the final human signature has to represent independent judgment rather than ratification.

Follow the recommendation

Imagine a hospital uses AI to suggest the next step in a patient’s care. Follow the suggestion to the person who may act on it. What evidence can that person inspect? Is there time to check it? Can that person reject it or call for help? Then identify who chose the system and who can suspend its use. This exercise maps responsibility across the whole process.

Accountability Verdict

Medicine should delegate aggressively where the evidence, boundary, and architecture support it. Refusing a safer or more effective tool merely to preserve a traditional human task would confuse accountability with nostalgia.

Medicine should also resist the opposite shortcut. Accuracy alone confers none of the institutional, fiduciary, professional, familial, or moral status that surrounds a medical decision. It cannot explain why a boundary was chosen, who authorized deployment, who can stop the system, or who must answer when the result harms a patient.

The practical rule is to govern the clinical chain one decision at a time. Give machines the work they can perform well. Give bounded autonomous systems authority only inside validated roles with escalation and monitoring. Require real human review when review is the safeguard. Preserve an independent human signature for Tier 3 and Reserved decisions where legitimacy requires it. Then make the developer, institution, clinician, and other actors answer for the parts of the system each controls.

The patient should never have to discover after something goes wrong that everyone involved believed someone else was responsible.

Chapter Eleven moves the same problem into a classroom, where the immediate stakes are usually lower but the long-term stakes may be just as profound. When every student can have an artificial tutor, the question becomes which parts of teaching can be delegated without quietly delegating the development of judgment itself.

Sources and Notes

1. Ethan Goh, Robert Gallo, Jason Hom, et al., “Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial,” JAMA Network Open, October 28, 2024, 7(10):e2440969. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2825395. Accessed September 10, 2026. Fifty physicians were randomized to conventional resources or conventional resources plus GPT-4 through ChatGPT Plus. Median diagnostic-reasoning scores were 76% versus 74%, an adjusted difference of 2 percentage points that was not statistically significant; the LLM alone scored a median 92% in an exploratory comparison. The paper notes that participating physicians had limited experience and little formal training with the tool. The study used an OpenAI product. OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH10-S1. Archive reference: pending.

2. Khaled Saab, Chunjong Park, Tim Strother, et al., “Advancing conversational diagnostic AI with multimodal reasoning,” Nature Medicine 32, 1726-1736 (2026), published May 14, 2026. https://www.nature.com/articles/s41591-026-04371-0. Accessed September 10, 2026. The randomized, blinded exploratory study used 105 simulated telehealth scenarios, compared multimodal AMIE with 19 board-certified primary-care physicians, and used 18 specialists as raters. AMIE scored higher on diagnostic accuracy and on 29 of 32 evaluation axes. The study was a simulation rather than a real-world clinical trial. It was funded by Google LLC or a subsidiary; the authors were affiliated with Google DeepMind or Google Research and disclosed Google employment and potential stock ownership. It is used as a vendor-developed capability signal, not as evidence of safe autonomous clinical deployment. Source record: DA-CH10-S2. Archive reference: pending.

3. U.S. Food and Drug Administration, “Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff,” final guidance issued January 6, 2026 and reissued January 29, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software. Accessed September 10, 2026. The guidance supersedes the 2022 version. The 2026 revision added enforcement discretion for certain software functions that provide a single clinically appropriate recommendation and moved the time-critical limitation into the independent-review analysis. Criterion 4 continues to require that the healthcare professional be able to independently review the basis for a recommendation rather than rely primarily on it. The final guidance explicitly retains its discussion of automation bias, including errors of commission and omission and the heightened concern when urgent action limits review time. FDA guidance states the agency’s current thinking and does not itself establish legally enforceable responsibilities. Source record: DA-CH10-S3. Archive reference: pending.

4. Ariel Leong, Risa M. Wolf, Roomasa Channa, et al., “Autonomous AI-assisted diabetic retinopathy screening at primary care is associated with increased presentation to eye care by at risk patients,” npj Digital Medicine 9, 310 (2026), published March 5, 2026. https://www.nature.com/articles/s41746-026-02460-5. Accessed September 10, 2026. The study identified 3,745 adults with diabetes referred to the Wilmer Eye Institute through ordinary provider referral or after a positive or non-diagnostic autonomous AI screening result. The authors reported an association with increased specialist presentation among African-American patients and explicitly cautioned that the exploratory observational analysis could not establish causation. Author Michael D. Abramoff disclosed patents and roles including investor, director, and consultant at Digital Diagnostics, the developer of the system. Source record: DA-CH10-S4. Archive reference: pending.

5. U.S. Food and Drug Administration, De Novo Classification Request DEN180001, IDx-DR, decision granted April 11, 2018. https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN180001. Accessed September 10, 2026. FDA classified the device as a retinal diagnostic software device and described its prescription indication as automatic detection of more-than-mild diabetic retinopathy in specified adults with diabetes using designated retinal imaging. The source is used to establish the narrow regulatory boundary of the autonomous screening role, not to generalize the device’s performance beyond its indication. Source record: DA-CH10-S5. Archive reference: pending.

6. U.S. Food and Drug Administration, “Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback,” August 18, 2026. https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request. Accessed September 10, 2026. The paper seeks feedback on risk assessment, a possible competency-based approach to premarket evaluation, postmarket monitoring, foundation models, agentic AI systems, and related issues. FDA requested comments by October 19, 2026 under docket FDA-2026-N-7874. FDA states that the paper is for discussion only, is neither draft nor final guidance, and does not itself propose or implement policy changes. Deadline verification accessed October 20, 2026. Source record: DA-CH10-S6. Update 107. Archive reference: pending.

7. American Medical Association, “AMA policies to ensure AI supports-not replaces-physician judgment,” June 10, 2026; “Aug. 28, 2026: State Advocacy Update,” August 28, 2026; and Washington Medical Commission, June 30, 2026 Interested Parties Meeting Packet, proposed interpretive statement “Clarifying eligibility for a license to practice medicine.” https://www.ama-assn.org/press-center/ama-press-releases/ama-policies-ensure-ai-supports-not-replaces-physician-judgment, https://www.ama-assn.org/health-care-advocacy/advocacy-update/aug-28-2026-state-advocacy-update, and https://wmc.wa.gov/sites/default/files/public/6.30.2026%20PIP%20Meeting%20Packet.pdf. Accessed September 10, 2026. The Washington Medical Commission proposed reading existing state law to permit only natural persons to hold a medical license, and a companion policy proposal stated that AI may inform but may not independently and autonomously assess, diagnose, treat, or prescribe medications. The AMA supported the licensure reading while objecting to placing full responsibility on a physician when an AI tool was embedded or mandated by an employer, health system, or payer or when a defect was not reasonably detectable. The Washington materials were proposals at the time of access, not final statewide rules. These are regulatory and professional-policy positions, not statements of settled national law. Source record: DA-CH10-S7. Archive reference: pending.

8. American Academy of Pediatrics, Srinivasan Suresh, R. Brandon Hunter, Barbara Periard, Juan Espinoza, Council on Clinical Information Technology, and Section on Innovation in Therapeutics and Technology, “Recommendations for the Development and Implementation of Generative Artificial Intelligence Tools in Pediatric Clinical Care: Policy Statement,” Pediatrics, published online October 3, 2026, e2026079037. https://doi.org/10.1542/peds.2026-079037. Accessed October 4, 2026. Professional policy guidance, not a clinical trial or binding law. Used for pediatric-specific evaluation, continuing monitoring, human oversight, and limitations of adult patient data. The task-specific delegation boundary is the author’s analysis. Source record: DA-CH10-S8. Update 067. Archive reference: pending.

9. Utah Office of Artificial Intelligence Policy and Division of Professional Licensing, agreement with Magic Health, Inc. and Magic Health Services, LLC, doing business as Nolla Health, signed September 22, 2026; “Authorized AI Pilots,” active demonstration October 5, 2026–October 5, 2027. https://commerce.utah.gov/wp-content/uploads/2026/10/RMA-Nolla-Health.pdf and https://commerce.utah.gov/ai/regulatory-relief-4/authorized-pilots/ Accessed October 6, 2026. Agreement Sections 3(E), 12, 19 and 20; Schedule B, Section 4.11. Narrow temporary regulatory mitigation, not state endorsement, certification, or proof of clinical safety. Used for eligible adults, topical-only scope, dual independent initial review, minimum stage duration and patient counts, weekly retrospective review, monthly sampling, escalation, written permission to advance, and termination authority. The interpretation about graduated delegation is the author’s analysis. Source record: DA-CH10-S9. Update 070. Archive reference: pending.

10. Medicines and Healthcare products Regulatory Agency and UK government, “Government backs recommendations of NHS doctors-led AI Commission,” October 6, 2026. https://www.gov.uk/government/news/government-backs-recommendations-of-nhs-doctors-led-ai-commission Accessed October 7, 2026. Official government response announcement, not an implemented law or clinical efficacy study. Used for acceptance of 44 recommendations, lifecycle assessment and monitoring, and AI Airlock Phase 3’s post-market surveillance focus. Planned guidance and roadmap remain prospective. Institutional monitoring, review triggers and suspension duties in the second paragraph are the author’s analysis. Source record: DA-CH10-S10. Update 073. Archive reference: pending.

11. Mahmud Omar, Reem Agbareia, Jolion McGreevy, et al., “Evaluating large language model responses to unsafe clinical instructions,” Communications Medicine, published September26,2026, accepted manuscript. https://doi.org/10.1038/s43856-026-01933-8. Mount Sinai, “Mount Sinai Study Finds Safety Prompts Can Help AI Models Make Safer Clinical Choices,” October8,2026. https://www.mountsinai.org/about/newsroom/2026/mount-sinai-study-finds-safety-prompts-can-help-ai-models-make-safer-clinical-choices. Accessed October9,2026. Peer-reviewed proof-of-concept evaluation of20 open-weight models,10,096,800 outputs,501 synthetic instances and100 MIMIC-IV-derived cases. Harmful-option selection is not a patient harm rate. Models and forced-choice tasks limit deployment generalization. The article predates the review window; October8 is the institutional announcement date. Evaluation recommendation is author analysis. Source record: DA-CH10-S11. Update 078. Archive reference: pending.

12. Anthony Abou Mrad, Jayanth S. Pratap, Aviona Conti, et al., “Stress-testing consentAI: a pipeline for evaluating a conversational educational adjunct to surgical informed consent,” npj Digital Surgery 1, article 27, published October 14, 2026. https://doi.org/10.1038/s44484-026-00025-3. Accessed October 15, 2026. Development-team study of 50 synthetic text conversations about one procedure and a study-specific rubric, not independent clinical validation. Half the sessions had only one clinician rater; clinician disagreement and the rubric’s limited validation constrain interpretation. Agreement is distinguished from failure detection; no tested architecture met the authors’ triage standard. J.K. Jopling discloses clinical-AI equity and robotics advisory interests; the paper states those companies had no involvement in this study. OpenAI systems are used in this Living Edition’s production workflow. Anthropic systems are used in the external critique workflow; see the transparency statement. The recommendation is author analysis. Source record: DA-CH10-S12. Update 097. Archive reference: pending.

13. Nigam Shah and Jonathan Chen, "Measuring What Medical AI Can Actually Do," Stanford Institute for Human-Centered Artificial Intelligence, October 19, 2026. https://hai.stanford.edu/news/measuring-what-medical-ai-can-actually-do. Ruoqi Liu et al., "PhysicianBench: Evaluating LLM Agents in Real-World EHR Environments," arXiv:2605.02240, first submitted May 4, 2026, https://arxiv.org/abs/2605.02240. Accessed October 20, 2026. The October 19 item is the researchers' explanation of prior benchmark work, not a new clinical trial. Controlled EHR evaluations do not establish deployment-wide failure or patient-harm rates. The Stanford explanation describes reconstructed records and the missed-order failure; the paper abstract corroborates execution-grounded, multistep evaluation but differs from the explanation in its reported model count. No model-count comparison is used. The second paragraph is the author's institutional analysis. OpenAI and Anthropic systems appear in the evaluations and are used in this Living Edition's production and external critique workflows respectively, under the standing disclosure. Source record: DA-CH10-S13. Update 105. Archive reference: pending.

14. Harsha Nori, Mayank Daswani, Christopher Kelly, et al., "Sequential diagnosis with language models," Science, October 22, 2026. https://doi.org/10.1126/science.aeh7930 Supporting primary account: MAI Health Team, "Building Medical Superintelligence," October 22, 2026, https://microsoft.ai/news/building-medical-superintelligence/ ; public dataset documentation, MIMIC-IV-Ext-SDBench v1.0.0, October 22, 2026, https://physionet.org/content/mimic-iv-ext-sdbench/1.0.0/ Accessed October 23, 2026. Developer-associated benchmark, not a patient-outcome trial. The comparison specifies physicians with reference resources. Full Science text was inaccessible; numerical claims were checked in the primary developer account, and methods/publication metadata against public dataset documentation. OpenAI systems are used in this Living Edition's production workflow. Second paragraph is author analysis. Source record: DA-CH10-S14. Update 115. Archive reference: pending.

15. Allison Kwon and Maya Sandalow, Bipartisan Policy Center, “Paying for Clinical AI: Principles for Policymakers,” October26,2026. https://bipartisanpolicy.org/article/paying-for-clinical-ai-principles-for-policymakers/ Accessed October27,2026. Primary policy analysis informed by two2026 roundtables; the authors explicitly say the principles do not represent advisory-board consensus. Used for payment incentives, demonstrated value and implementation costs. Recommendations are distinct from enacted payment rules and clinical outcomes. The second paragraph, including the specific budget for review, incident investigation and suspension, is the author’s institutional analysis. Source record: DA-CH10-S15. Update 119. Archive reference: pending.

16. Elizabeth M. Norell et al., “Generative Artificial Intelligence and Maternal Health Outcomes in Pregnant and Postpartum Individuals: A Randomized Clinical Trial,” JAMA Network Open 9(11):e2641062, November4,2026. https://doi.org/10.1001/jamanetworkopen.2026.41062 Accessed November5,2026. Randomized trial; self-reported outcomes, active comparator and low baseline symptoms constrain interpretation. Secondary/exploratory findings do not establish harm. Second paragraph is author analysis. Source record: DA-CH10-S16. Update 132. Archive reference: pending.