The ACP has issued ethical guideposts for AI at the bedside. There are three.
Relationality, self-governance, competence. The position paper's starting premise is that consensus on privacy, disclosure and fairness has not been reached, and clinicians need guidance anyway.
Most published guidance on artificial intelligence in medicine addresses institutions: how a health system should procure a model, validate it, monitor it, govern it. Very little of it addresses the clinician holding the output while a patient waits. A position paper from the American College of Physicians, published 31 August, is explicitly aimed at that gap [s1].
The premise
The paper opens with an observation about pace rather than about technology. Artificial intelligence in health care is being rapidly adopted, in what the authors describe as perhaps one of the most rapid implementations of technology in medicine to date [s1]. Many guidelines for its use have been issued, but consensus has not been achieved on privacy, disclosure and fairness [s1].
That is an unusual thing for a professional society to write down. It concedes that the governance layer has not settled the foundational questions, and argues that patients and physicians need ethical guidance at the point of care regardless [s1].
The paper also makes a terminological move worth flagging, because it recurs through the document: AI in this context is better thought of as augmented intelligence — a tool that provides assistance to physicians and patients [s1]. That framing is doing normative work, not just descriptive work. It positions the technology as an input to a clinician's judgement rather than a substitute for it, and everything that follows is built on that positioning.
The three guideposts
The paper's contribution is three guideposts for the ethical use of AI, all rooted in the patient-physician relationship: relationality, self-governance, and competence [s1]. The paper describes these as informed by principles of medical ethics, clinical integrity, and physician-independent practical reasoning [s1].
The terms are chosen carefully, and it is worth being clear that the abstract names them without defining them — the definitions live in the full position paper, and any gloss offered here would be inference rather than reporting. What can be said from the framing alone is where the paper locates the ethical question. Not in the model, and not in the procurement process, but in a relationship between two people that the model has entered.
"Competence" is the term that carries the most weight, particularly alongside its stated foundation of physician-independent practical reasoning [s1]. If a clinician's judgement is increasingly exercised in the presence of a model's recommendation, whether that judgement remains independent is not a rhetorical question.
The same journal published on the same day an Annals On Call discussion titled "Is Artificial Intelligence Eroding Clinical Competence?" [s2]. The pairing indicates the question is being treated as live rather than settled — though that item is a recorded discussion, not a study, and no evidence about competence erosion is presented in either.
What a position paper is and is not
This is a normative document from a professional society. It contains no trial, no cohort, no measurement. It does not establish that AI use degrades clinical skill, does not quantify any harm, and does not evaluate any specific product [s1]. Position papers of this kind carry influence through adoption — by training programmes, by credentialing bodies, by malpractice standards over time — rather than through evidence.
Its practical value lies in what it makes askable. A framework organised around relationality, self-governance and competence generates specific questions a clinician can put to a deployment: what is the patient being told; who set the standard for acceptable use; and whether the arrangement preserves independent reasoning or quietly outsources it [s1].
Where it sits
The paper's own account of the field is that many guidelines have been issued and consensus has not been reached [s1]. What distinguishes this document from most of them is its unit of analysis: it is written for the individual physician and the individual consultation rather than for the institution doing the deploying.
What it does not resolve is disclosure, which the paper itself names as unsettled [s1]. Whether a patient must be told that a model contributed to a recommendation, and in what terms, remains the open question with the most direct consequences for consultations happening now.
What to watch is whether the three guideposts are picked up in operational documents — training curricula, health-system policies, professional standards — or remain a statement of principle. That is the difference between a framework and a position.
Sources
- [s1] Ethics and Professionalism in Artificial Intelligence and Medical Practice: A Position Paper From the American College of Physicians. Annals of Internal Medicine, 31 August 2026. https://doi.org/10.7326/ANNALS-26-01792
- [s2] Annals On Call - Is Artificial Intelligence Eroding Clinical Competence? Annals of Internal Medicine, 31 August 2026. https://doi.org/10.7326/ANNALS-26-03397-OC
Sources
- Ethics and Professionalism in Artificial Intelligence and Medical Practice: A Position Paper From the American College of Physicians — Annals of Internal Medicine , August 31, 2026
- Annals On Call - Is Artificial Intelligence Eroding Clinical Competence? — Annals of Internal Medicine , August 31, 2026
Sepsis AI reaches an AUROC of 0.88 and a positive predictive value of 34.2%
A network meta-analysis of 53 studies and more than 7 million admissions finds machine learning out-discriminates traditional sepsis scores — and would raise roughly two false alarms for every real one.
UAE hospitals are using AI to flag sepsis six hours before doctors would catch it
The tool reads vital signs, lab results, and history the way clinicians already do — just faster and continuously. Its accuracy tops out around 90%, which means it's also still wrong a meaningful share of the time.
An Israeli AI cut antibiotic mismatch by 30%. A third of doctors ignored it anyway
Maccabi Healthcare Services studied 626 of its own physicians to work out who follows algorithmic prescribing advice — and found the pattern was about practice structure, not just attitude.
Clinical software that gives one answer is no longer a device, the FDA says
Revised guidance reverses the 2022 position that a single recommendation made software a device. The agency's own criterion — that a clinician can independently review the basis — now carries the weight.