The quiet gap in physicians’ skill-loss worry A 2026 American Medical Association survey of 1,692 US physicians found that 88%, or nearly nine in ten, are at least somewhat concerned that AI could contribute to a loss of clinical skills.
That number sounds straightforward, but there is an interesting gap underneath it.
Physicians are much more concerned about skill loss among the next generation than about their own skills. 70% are concerned that today’s medical students and residents could experience skill loss, while only 28% are concerned about their own skills declining. The difference matters. Physicians appear to see the risk more clearly for people who are still developing their clinical judgement. They may have spent years building their own skills before AI became part of everyday clinical work. Students and early-career doctors may have more of that learning replaced by tools that provide answers before they have worked through the problem themselves.
That makes skill loss less about whether AI can perform a task and more about what happens when a person no longer needs to practise that task regularly.
Who worries most about their own skills
By experience:
Early career physicians who have been in practice for fewer than 10 years; 35%
Late career physicians who have been in practice for over 30 years; 21%
By specialty:
Primary care physicians; 34%
physicians in medical specialties; 25%
physicians in surgical specialties; 24%
The pattern is notable. The highest level of concern is not necessarily among the specialties where AI might appear most technically advanced. It is higher among physicians who are earlier in their careers and among those working in primary care. One possible reason is that these groups rely heavily on broad clinical reasoning. Early-career physicians are still developing their clinical judgement. Primary care doctors also need to consider a wide range of symptoms, conditions and possible explanations rather than focusing on one narrow area.
A second worry, from a different health system
The study found that 30.6% of Swedish physicians and 33.2% of nurses said AI use would reduce their sense of meaning in their work. This is related to skill loss, but it is not the same concern. The AMA figures ask, in effect, whether physicians might become less capable at something they currently do. The Swedish findings raise a different question: if AI takes over more of the work, will the job still feel like the same profession?
That distinction is important. People can welcome automation and still be concerned about what is lost when parts of their work are removed.
Physicians want the automation anyway
The numbers suggest that physicians are not rejecting AI. They want it to reduce the burden of work, particularly tasks that are repetitive or administrative and take time away from patients. The concern is what happens when automation goes beyond reducing that burden and starts taking over the judgement that clinicians need to exercise themselves.
The distinction is simple; physicians want the work to become easier, but they do not want the reasoning to disappear. AI can take care of some of the workload while leaving clinicians to interpret the information, consider the options and make the final judgement.
The design point that explains it
A different approach is to use AI to surface relevant information, highlight possible conditions or provide suggestions, while leaving the clinician to work through what that information means for the patient. The tool can reduce the time spent finding and processing information without taking away the need to interpret it.
It is similar to using a calculator. A calculator that only gives you the final number does the work for you. One that shows how the calculation was reached still helps you understand and check the process. In the same way, clinical AI can save time while keeping the clinician involved in reasoning.
The important point is not simply whether AI makes clinical work faster. It is whether it saves time without removing the practice that helps clinicians maintain their judgement.
Where Autoderm fits This distinction is relevant to how dermatology AI can be used in a clinical workflow.
Autoderm is designed to return informational condition suggestions for a clinician to review and reason through, rather than providing a finished decision for the clinician to simply accept. The clinician remains responsible for interpreting the suggestions in the context of the patient and the wider clinical information.
Autoderm is a dermatology API, rather than a standalone clinical decision-maker. Its informational condition suggestions are intended to support the clinician’s reasoning process. The available evidence reports documented suggestion accuracy in the range of 72–93% across published studies, with the higher figure coming from a comparative validation against a five-condition differential in the Coachella Study (2025). Autoderm also holds CE marking as a legacy Class I device under the MDD, with MDR Class IIa in progress. Further information on its clinical applications and evidence can be reviewed on the Autoderm solutions pageAutoderm clinical evidence page
The takeaway The real question for a clinical AI tool is not simply whether it saves time. Most useful tools will save time in some way. The more important question is what happens to the reasoning that sits behind the work. If AI can reduce the burden while still leaving clinicians to interpret, question and decide, it can support the work without taking away the practice that keeps those skills alive.
Physicians already appear to recognise that trade-off. The way clinical AI is designed will help determine whether it reduces work while keeping clinicians thinking.
References
American Medical Association. (2026). 2026 physician survey on augmented intelligence. AMA Center for Digital Health and AI. https://www.ama-assn.org/system/files/physician-ai-sentiment-report.pdf
Blease, C., et al. (2026). Generative AI use by physicians and nurses in Sweden: Development of a national cross-sectional survey [Preprint]. JMIR Preprints. https://preprints.jmir.org/preprint/99788
