Learners Brought Their Own Models to the Wards

Key Takeaways

• Learners already use consumer AI during clinical work, so accredited activities compete with those tools.
• The stronger CME response is not a ban or a cloned question bank; it is supervised critique of AI output.
• Do not market personalized AI learning unless the human review process is real, visible, and role-specific.

Students and residents are already using ChatGPT, OpenEvidence, and Gemini on the wards. Learners can get instant answers elsewhere, so accredited education has to offer something those tools do not.

The CME offer is critique, not another chatbot

Vinay Prasad’s education comments treat the ward use as a given: learners are already using ChatGPT, OpenEvidence, and Gemini while seeing patients. Banning the tools misses the harder task—teaching people to ask better questions, test recommendations, and defend the clinical path they choose.

Accredited education is no longer competing only with other CME. It is competing with the answer engine in the learner’s pocket. In an Oncology Data Advisor interview, an oncology education founder described board-prep tools as 10-15 years behind the literature and proposed AI-supported paths for fellows, attendings, PAs, and NPs, with clinicians deciding what is released to learners. That interview is product-linked and should not be treated as market consensus. It still names a real provider problem: if the accredited product looks like delayed board prep, learners will route around it.

Rochelle Walensky added the literacy limit. In a JAMA+ AI Conversations episode, she said, "We know they're good at taking board exams." Models can win knowledge tasks while still carrying training-data bias, reinforcing user assumptions, and struggling with bedside human work. That is why an earlier brief treated AI literacy as failure drills rather than feature tours: the learning value is in finding where the tool is wrong, incomplete, overconfident, or too agreeable.

A generic AI overview is too shallow, and an unreviewed chatbot is too risky. The activity that still has a reason to exist is one where the learner critiques an AI-generated answer, identifies bias or missing evidence, compares it with a vetted standard, and justifies the next step. If the learner already uses AI on the ward, what does the activity add that the model will not?

What CME Providers Should Do Now

• Name which learner roles—student or resident, fellow, attending, APP—get which AI-supported path. Do not design one generic chatbot for everyone.
• Replace at least one ban-or-overview segment with critique of a biased, outdated, or overly agreeable AI output.
• Make the human review process visible to learners before advertising AI personalization.
• Ask whether the activity still has value if the learner already uses OpenEvidence or ChatGPT during clinical work.

What CME teams should reconsider

A ban, a delayed question bank, and an unreviewed generator all lose to tools already on the ward. What still holds up is human-reviewed content, role-specific paths, and repeated practice challenging the model before trusting it.