Small Peer Networks Outperform Star Faculty in Changing Clinical Behavior

Physician groups that shared information made fewer diagnostic and treatment errors than physicians working alone, according to a social-network keynote summarized this week. The evidence is narrow—one CME educator podcast recap of conference keynotes—but the implication for providers is concrete: learning design has to create repeated peer testing, not just deliver expert interpretation.

Peer interaction is the intervention, not the add-on

The strongest signal came from Damon Centola’s keynote on social networks, summarized by Write Medicine. The recap described a study in which physician groups working through medical scenarios made significantly fewer errors when they shared information and worked together than when physicians worked alone.

The point for CME providers is not simply “add discussion.” It is that behavior change may depend on multiple reinforcing interactions across a network. That challenges the default conference model in which a high-profile faculty member sits at the center and change is expected to diffuse outward.

This sharpens a thread we saw in an earlier brief on clinicians routing around long CME and repeat training: clinicians value formats that let them process judgment with peers, not just consume polished expertise. This week’s signal adds a mechanism—complex contagion—and a measurable outcome: fewer errors in structured physician groups.

For CME teams, the question is whether flagship programs are built to create those reinforcing interactions. A lecture followed by a short Q&A is not the same as a cohort, cross-departmental case exchange, or peer-feedback loop that continues after the event.

AI education needs bias checks built into the format

The same Write Medicine episode summarized Immani Shephard’s keynote on AI in healthcare education. The concern was not only that AI can be wrong, but that it can amplify biased data and then reinforce clinician bias through human-AI feedback loops.

The example in the recap was the flawed race-based GFR assumption, linked to delayed interventions, prolonged dialysis, and reduced transplant opportunities. The broader design implication is cross-specialty: AI literacy cannot stop at prompt-writing, tool familiarity, or productivity use cases.

CME activities that include AI outputs should make skepticism visible. Learners need planned moments to ask: What data may be missing? Whose outcomes might worsen? Are social determinants of health being considered? When should AI augment judgment, and when should the clinician pause?

For providers, this changes faculty preparation and outcomes planning. If an activity uses AI, the evaluation should not only ask whether learners understood the tool. It should ask whether they can identify bias risk and apply the tool equitably in clinical decision-making.

What to audit now

• Calculate how much time in major live programs is spent in structured peer interaction versus expert delivery.
• Pilot one cohort-based program with deliberate cross-site or cross-department connections, then measure practice change at 3–6 months.
• Add bias-reflection prompts and equity checkpoints to any activity that demonstrates or relies on AI outputs.

What CME teams should reconsider

The week’s quieter signal is that format is carrying more of the educational burden. Peer networks and bias checks point to the same issue: clinicians need structured ways to slow down, compare interpretations, and test judgment before practice changes. The question for providers is no longer whether interaction belongs in the agenda. It is whether the agenda gives interaction enough structure, repetition, and measurement to change what clinicians do after the session ends.