Teach AI Without Shortchanging Clinical Reasoning

Key Takeaways

• AI education should not be measured only by adoption or speed; CME teams need objectives that protect clinical reasoning.
• AI-assisted assets need editorial judgment. Clinicians are noticing when slides and summaries lose human narrative.
• Conference follow-ons should help learners sort implementation choices, not simply compress meeting volume.

Clinician discussion this week framed AI in medical education as a reasoning problem, not just a tool-adoption problem. The examples were oncology-led, but the CME implication is broader: learners need help knowing when AI can accelerate work and when it may weaken the learning process.

AI training needs a reasoning floor

The strongest clinician signal came from heme-onc education. A physician thread highlighted new research on fellows’ AI use and interest, arguing that AI is already part of medicine and that best practices need to be taught soon, not left to informal habit formation (Ariela Marshall MD). A separate fellowship-focused episode from Two Onc Docs made the educational risk more concrete: tools such as Open Evidence may be useful for retrieval, but first-year fellows still need to work through resources such as NCCN algorithms themselves rather than jump straight to a quick answer (Two Onc Docs).

For CME providers, that changes the center of gravity. AI modules cannot stop at prompts, disclosure, or verification. They need to state which parts of the learning task must remain learner-owned: forming a differential, tracing an algorithm, explaining why an option is inappropriate, and recognizing when a tool’s answer is too shallow for the case.

The second part of the signal was editorial, not technical. One oncology clinician described fatigue with AI-created slides, blogs, notes, brochures, posters, papers, and reviews, while also saying, “Yes, I use AI daily because it is very useful for search and task management.” The complaint was not anti-AI; it was about low-value output. The same post ended with the sharper standard: “I want to read something with novelty.” (Yu Fujiwara, MD)

For CME teams, the question is simple: where does an activity require AI help, and where does it require a human educator’s judgment, narrative, and explanation?

Meeting coverage should become curriculum, not compression

The ASCO26-adjacent material this week pointed in the same direction from a different angle. Much of this evidence came from society, academic, or organization-led channels, so it is best read as an emerging provider signal rather than broad independent clinician consensus.

The relevant point is not another list of oncology abstracts. In post-meeting discussion, ASCO was described as a condensed setting where general oncologists can get current information across a wide disease range (CancerCast). Other ASCO26 analysis emphasized careful interpretation of endpoints, magnitude of benefit, and therapeutic burden rather than passive acceptance of “practice-changing” framing (ecancer). In colorectal cancer coverage, faculty discussion moved quickly from trial data into how clinicians vary in real-world use of ctDNA MRD testing outside trials (Research To Practice).

That is the provider opportunity: conference follow-ons should help learners turn meeting volume into structured decisions. We saw a related pattern in an earlier brief on ASCO26’s MedEd signals, where trainee curricula and implementation support mattered more than abstract-by-abstract recaps.

For CME teams, the concrete question is whether a conference product teaches learners how to decide, explain, and implement—or merely helps them remember what was presented.

Questions for CME Teams This Week

• Audit AI education for explicit foundational-reasoning objectives, not only tool-use objectives.
• Label where AI use is encouraged, discouraged, or compared against manual guideline review inside the activity.
• Add at least one assessment step that requires learners to explain their reasoning before seeing an AI-assisted answer.
• Build conference follow-ons around vignettes and implementation choices, not only “top abstracts” summaries.

The line to draw

The useful distinction is not AI versus no AI. It is retrieval versus reasoning, generated content versus educator narrative, and meeting coverage versus curriculum. CME providers that make those distinctions visible will be better positioned than providers that treat AI speed and conference volume as learning outcomes by themselves.