AMA Physician AI Sentiment Report (2026)

The AMA Physician AI Sentiment Report is the third wave of a longitudinal survey of U.S. physicians conducted by the American Medical Association's Center for Digital Health and AI. It was published in March 2026 and fielded January 15 - February 2, 2026, with a sample of 1,692 physicians. The prior waves were conducted in 2023 (n=1,081) and 2024 (n=1,183). The report measures physician awareness, adoption, sentiment, and regulatory preferences regarding AI in clinical practice; it provides longitudinal sentiment and adoption data for a high-skill clinical workforce outside the software- and customer-support-sector evidence found in sources such as Generative AI at Work, Generative AI and the Nature of Work, and Stanford HAI AI Index Report 2026.

Adoption and use

Physician-reported AI use rose from 38% (2023) to 66% (2024, a figure that includes uncertain awareness) to 81% (2026). Among 2026 respondents, 72% incorporate one or more use cases, 9% are uncertain, and 19% report no use. Use-case density also grew, from an average of 1.1 use cases per physician in 2023 to 2.3 in 2026. Personal AI use outside work is widespread: 77% of physicians use AI at least weekly and 35% use it daily.

The most common use cases are documentation-heavy: research and standards-of-care summaries (39%, up 26 percentage points versus 2024), discharge instructions and care plans (30%), chart documentation and billing (28%), and chart summaries (28%). Assistive diagnosis trails at 17%.

Sentiment

The share of physicians describing AI as an "advantage to patient care" rose from 65% (2023) to 69% (2024) to 76% (2026). The share who are "more excited than concerned" rose from 30% to 35% to 37%, while 40% are equally excited and concerned. Across measured dimensions, the only one with a net-harmful expectation is patient privacy (13% helpful versus 41% harmful). Work efficiency (78% helpful versus 7% harmful), diagnostic ability (74% versus 11%), and stress and burnout (64% helpful) are net-positive.

Burnout

Industry discourse frequently cites AI as a burnout intervention, and the survey provides aggregate physician expectations on this point. 70% of physicians expect AI to offload or replace clinical tasks contributing to burnout, and 73% expect automation of administrative workload. Only 13% are concerned that AI will worsen stress or burnout. Helpful-for-burnout sentiment rose from 48% (2024) to 64% (2026). These figures are physician expectations of impact rather than measured burnout outcomes (see Caveats).

Skill loss

88% of physicians are very, somewhat, or mildly concerned about skill loss overall. Concern is distributed unevenly across whose skills are at issue (ratings of 4 or 5 out of 5): 70% are very or somewhat concerned about current medical students and residents (with 46% rating concern at 5), 40% about colleagues, and 28% about themselves. Personal skill-loss concern is highest among early-career physicians with fewer than 10 years of practice (35%), versus 24% among late-career physicians with more than 30 years. By specialty, concern runs higher in primary care (34%) than in medical specialties (25%) or surgical specialties (21%).

The report frames concern as concentrated among physicians whose skills are still being built. The AMA survey functions as a self-perception counterpart to the behavioral findings of the Lancet endoscopist deskilling study, which documented adenoma detection rates falling from 28.4% to 22.4% after AI exposure; the AMA data captures physician self-perception of skill erosion (88% concerned), aligning with the practice-through-exposure mechanism that Budzyn et al. documented behaviorally.

Patient use of AI

The 2026 wave introduced new measures of physician perceptions of patient AI use. 30% of physicians believe a majority of their patients are probably using AI for health information; 8% say a majority disclose it; and 29% have had zero disclosures. 70% say patient use of general-purpose chatbots for health information is positive or neutral. Physicians report comfort with patients using AI for medication questions (68%), general health (64%), and personalized questions (55%). Comfort drops sharply for interpretation tasks: nearly half would never or rarely want patients using AI for pathology (49% oppose) or radiology (46% oppose).

Regulatory and adoption priorities

The survey records ranked physician priorities for AI regulation, relevant to AI LEAD Act (S. 2937), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment), and the EU AI Act's high-risk medical device provisions. Among trust-builders (percentage rating an item 4-5 in importance), validated safety and efficacy plus ongoing monitoring rank highest at 88%, followed by data privacy assured by the institution and EHR vendor at 86%. Concern about non-institutional tool privacy (71%) exceeds concern about institutional tools (42%).

Physician regulatory priorities, ranked by average rank, are:

1. Clear liability frameworks (average rank 3.0; the top choice for 31% of respondents)
2. Post-market surveillance (3.4)
3. Increased FDA oversight of AI-enabled medical devices (3.6)
4. Oversight of AI tools not classified as devices (4.1)
5. Payor use of AI in medical-necessity determinations (4.4)

On involvement and training, 85% want active involvement in adoption decisions (55% want to be consulted, 30% want to be responsible), and 92% want more AI training. Only 11% of those who received training report receiving "a lot of" training. Training exposure varies by practice setting: hospital (76%), group (71%), and solo (55%).

Relationships

• supports: AI and Productivity - healthcare-sector deployment-reality evidence (physician adoption trajectory and perceived efficiency gains)
• supports: AI Labor Disruption - high adoption plus documentation concentration indicates near-term task restructuring in medicine
• supports: Lancet Endoscopist Deskilling Study (2025) - corroborating subjective evidence for AI-induced skill erosion in clinical practice; the AMA survey is the physician self-perception signal, the Lancet study is the measured-endpoint signal
• related: AI Deskilling - if/when created as a concept page, this is foundational survey evidence
• related: Generative AI at Work - both find largest adoption and relief among workflow tasks (documentation here, ticket resolution there); similar leveler expectation pattern
• related: MIT NANDA — The GenAI Divide (State of AI in Business 2025) and other business-reality sources - medicine shows a faster, more positive adoption curve than the 95% enterprise pilot-failure finding in other sectors; worth comparing once ingested
• related: AI LEAD Act (S. 2937), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) - physician regulatory priorities (liability frameworks, FDA post-market surveillance) map onto these state and federal frameworks

Tensions and caveats

1. Self-report versus behavior. All findings are self-reported awareness and use; the survey does not independently verify organizational deployment. The 81% figure is physician-reported awareness, not necessarily integrated clinical use.
2. Instrument drift limits strict comparability. The 2024 wave added an "uncertain which tools my practice offers" response option; the 2026 wave added two new use cases (patient chatbots, virtual patient simulations) and included qualified partial responses for the first time. Some of the 38% to 81% growth reflects instrument sensitivity, though the magnitude of change is too large to attribute solely to methodology.
3. Burnout data is expected impact, not measured outcome. The 70% and 73% burnout-relief figures are physician expectations. A "burnout scores dropped 42% to 35% with ambient AI" data point cited in the source-recommendation pre-brief is not in this report's primary findings and likely references separate AMA ambient-scribe pilot data; the report itself measures sentiment, not realized burnout reduction. Flagged for follow-up.
4. Privacy paradox. Physicians cite privacy as the leading concern dimension (41% expect harm) and 71% worry about non-institutional tools, yet 77% use AI weekly in personal contexts. The survey does not resolve whether this reflects compartmentalization or genuine comfort with the personal-versus-clinical boundary.
5. Skill-loss self-versus-other asymmetry. Physicians are about 2.5 times more worried about trainees' skill loss (70%) than their own (28%). The report notes this is consistent with self-serving bias and also with the Lancet finding that deskilling effects concentrate among those learning the skill through AI-exposed practice.

Confidence notes

The survey methodology is transparent and large-N (1,692), with documented sampling and weighting details. Core adoption trends (use cases, comfort, burnout expectations) are high confidence as sentiment data. As predictors of actual clinical outcomes (whether AI reduced burnout or eroded skills), the report is low-to-medium confidence, because it measures beliefs rather than outcomes. For outcome data see Lancet Endoscopist Deskilling Study (2025) and future ambient-scribe randomized controlled trials.