Medical Schools Lack Consensus on Teaching AI Critical Thinking

Medical Schools Lack Consensus on Teaching AI Critical Thinking

Medical schools around the world have no agreed-upon framework for teaching artificial intelligence, leaving future physicians unprepared to critically evaluate AI-generated clinical guidance, according to research from Harvard Medical School.

Felipe Fregni, a professor at Harvard Medical School and the Harvard T.H. Chan School of Public Health, analyzed 54 studies from 22 countries and found zero consensus on how to integrate AI education into medical training. The findings were published in npj Digital Medicine and first reported by Medscape.

The gap is particularly concerning because AI systems produce authoritative-sounding text even when factually wrong, Fregni told Medscape's Portuguese edition. Students who haven't learned to distinguish fluency from accuracy will accept flawed machine output without question.

Why it matters

As AI tools become embedded in electronic health records and clinical workflows, physicians who can't critically appraise machine-generated recommendations will add diminishing value to patient care. The issue isn't cognitive laziness—it's that medical education hasn't systematically taught students to interrogate AI the way they would a research paper or differential diagnosis.

The curriculum gap

Fregni identified three stages medical schools should address. First, foundational AI literacy: how models work, their limitations, and ethical considerations. Harvard offers this as a pre-enrollment summer program because incoming students arrive with wildly different technical backgrounds.

Second, clinical integration during rotations—teaching students to use AI responsibly for diagnosis support and documentation while revisiting privacy and bias concerns in real hospital settings.

Third, and most challenging: teaching students to critique AI output. Fregni assigns exercises where students must identify errors, missing context, or fabricated references in machine-generated clinical text.

The obstacle isn't student resistance. "Often the teacher pretends the student is not using AI, and the student pretends not to be using it, when in fact everyone is using it," Fregni said. "So there is no learning."

Faculty training lags behind

The bigger barrier is faculty preparedness. Unlike established specialties where instructors share common language and methods, AI expertise varies wildly among medical school teachers. Many have never received formal training themselves.

Fregni recommends pairing clinicians with bioinformatics PhDs to deliver depth alongside practical application. Schools also need annual curriculum updates—difficult in regions like South America where education ministries must approve changes, slowing adaptation to rapidly evolving technology.

Productive difficulty versus fluency

Fregni disputes concerns about "cognitive complacency"—the fear that AI will make students think less. He compares it to calculators: researchers lost manual computation skills but gained sophisticated statistical interpretation abilities.

The real danger is that AI eliminates productive struggle. Effective learning requires retrieval practice, mistakes, and feedback—uncomfortable processes that feel less efficient than reading AI summaries. "What feels efficient generally produces little durable learning," Fregni explained.

He advocates for metacognition training: teaching students to calibrate their own confidence and recognize knowledge gaps. Assessments must test reasoning, not memorization. "Why did you choose this approach? What would make you change your mind?" are the questions exams should pose.

Without structured critical thinking education, Fregni warned, physicians will struggle to add value as AI capabilities expand over the next five years.

These findings were first reported by Medscape based on an interview with Fregni and his research published in npj Digital Medicine.