AI Literacy Needs Failure Drills, Not Feature Tours

Key Takeaways

• AI education is moving from awareness to competency: clinicians need to rehearse how to detect and override predictable failure modes.
• The teachable risks this week were specific: automation bias, sycophancy, hallucinated citations, anchoring, and workflow overdependence.
• For CME teams, the design question is whether AI activities measure verification behavior, not just whether learners understand the technology.

Clinicians are naming specific AI failure modes and demanding training that builds verification habits. A narrow but useful source mix—a journal podcast, institutional discussion, and independent clinician thread—shows the shift from general awareness to concrete guardrails.

AI literacy now needs a failure-mode curriculum

The clearest thread came from a JAMA+ AI Conversations episode that treated AI risk less as a general warning and more as a set of teachable behaviors. The discussion named automation bias: when a system is right often enough, human supervision becomes weaker because checking stops feeling worth the effort. It also named sycophancy and anchoring, including the practical warning: “And so realize if you want an objective answer, don't tell AI what you're thinking.”

That matters for CME because “AI literacy” can no longer mean a broad primer on what large language models are. The competency is whether clinicians can spot when a model is agreeing too readily, producing plausible but false citations, reinforcing a flawed first impression, or quietly taking over a workflow skill that clinicians still need when the system is unavailable. We saw a related pattern in an earlier brief on AI training that names failure modes; this week’s difference is that the guardrails are becoming more concrete and operational.

The medical-education signal pointed in the same direction. In an AAMC discussion of AI and admissions, speakers framed AI proficiency around critical thinking, ethical use, and not outsourcing judgment. One panelist also described work underway to develop AI competencies and a course for third- and fourth-year medical students on AI in healthcare. That is institutional education, not broad clinician consensus, but it reinforces the same design problem CME teams face with practicing clinicians: baseline familiarity is uneven, and the risks are tied to judgment under pressure.

The independent clinician signal was smaller but sharper. A physician post about AI in complex clinical decision-making emphasized that AI is moving beyond diagnosis while still carrying caveats around bias, confabulation, automation risk, and safety thresholds (source). The provider implication is broader: CME should help clinicians practice when to trust, when to verify, and when to override.

The concrete implication: audit AI-related education for moments where learners must identify a failure mode, check the model’s work, and decide what they would do differently in the clinical workflow.

What CME Providers Should Do Now

• Replace generic AI safety warnings with case prompts that force learners to identify automation bias, sycophancy, hallucination, or anchoring.
• Build short verification drills into AI activities: source-checking, second-opinion prompts, citation review, and “what would make you override the output?” decisions.
• Separate baseline AI literacy from role-specific application: every clinician needs failure-mode recognition, but workflows differ by specialty, setting, and responsibility.

What CME teams should reconsider

The weak point in many AI curricula is not enthusiasm or access. It is whether the education changes supervision behavior after the first impressive answer appears. If an activity does not rehearse calibration, verification, and override, it may be teaching comfort with AI without teaching control of it.