When Clinicians Disclose AI Use, Patient Trust Drops Faster Than Expected

Key Takeaways

• AI education now has to rehearse clinician disclosure, judgment, and workflow decisions—not just tool use.
• The strongest provider implication is observable behavior: what clinicians say to patients, when they check AI output, and when they avoid using it.
• Cost-effectiveness measurement in accredited CME is still mostly baseline work; teams should start small before making value claims.

Clinician conversation this week pointed to a harder requirement for AI education: clinicians need to explain, check, and justify AI use in the encounter. The AI examples came from radiology, oncology, and general medical discussion, but the provider implication is broader: CME has to make trust-preserving behaviors visible enough to teach and measure.

AI disclosure is becoming a communication skill

The sharpest signal came from a JAMA AI discussion of a survey of about 1,300 U.S. adults: when identical physician ads disclosed AI use, perceived competence, trustworthiness, empathy, and willingness to book an appointment dropped. The discussion did not argue for hiding AI use. It argued that clinicians need better language for explaining how AI supports care while making clear that the clinician remains responsible (JAMA+ AI Conversations).

That moves AI education out of the “tool demo” lane. A clinician who can operate an AI-enabled documentation or imaging tool may still be unprepared for the patient-facing moment: Why was AI used? What did it contribute? What did the clinician verify? What happens if the AI finding conflicts with the clinician’s judgment?

Workflow fit is the other half of the same problem. In a radiology leadership discussion, one clinician described impression generators as useful but not necessarily time-saving: “So far, we've found tools that our radiologists say, they may not make me much faster, if at all.” The more important benefit was lower cognitive load, paired with a warning that tools can create extra review work if they sit outside the normal workflow (AJR Radiology Trailblazers). A separate clinical AI conversation made a similar point from the other direction: AI can help with complex uncertainty, but in routine cases it can add alert fatigue, automation bias, and bloat (Healthcare Unfiltered).

We saw a related pattern in an earlier brief on AI adoption barriers: the problem is rarely awareness alone. This week’s update is more specific. CME teams should build AI cases around disclosure language, verification checkpoints, and “do not use AI here” decisions, then assess whether learners can actually perform those behaviors.

Cost claims need a starting line

The second signal was narrower and came from provider-owned educational content, so it should be treated as conceptual framing rather than a definitive benchmark. Still, the finding is useful for CME operators: a 10-year landscape review found that cost concepts are barely visible in accredited CME outcomes literature. Out of roughly 1,000 screened records, 32 mentioned cost in some way, and 13 reported an actual cost variable such as cost per participant, cost per unit of educational improvement, or modeled cost savings (JCEHP Emerging Best Practices in CPD).

The useful distinction is between the cost of the educational activity itself and the effect of that education on downstream healthcare costs. The latter is where the literature appears especially thin. That matters because health systems and accreditors may ask CME teams to speak in the language of value, but most providers do not yet have a stable internal method for doing so.

The implication is not to bolt a health economics model onto every activity. It is to start with a baseline: what did this activity cost to plan, deliver, and measure; what changed; and what denominator makes the comparison honest? For high-volume or system-partnered programs, even a simple cost-per-learner and cost-per-primary-outcome-change field would make future value conversations less anecdotal.

What CME Teams Should Do Now

• Add one AI disclosure prompt to relevant cases: what the clinician says, what they verified, and how they preserve responsibility for the decision.
• Map AI activities to workflow checkpoints: when the tool is useful, when it adds burden, and who confirms output before it reaches the record or patient.
• For one high-volume program, capture total activity cost, participant count, and the primary measured change before making any cost-effectiveness claim.

What to reconsider

If an AI activity ends with tool proficiency, it is probably stopping too early. If an outcomes report ends with confidence and intent, it may not be ready for value conversations. The week’s two signals point to the same operating discipline: teach the behavior clinicians must perform in public, and measure enough of the cost to know what the education actually required.