In emergency medicine, sequence is half the job. Airway before breathing. Check the glucose before you call it a stroke. I’ve come to believe AI at the bedside needs the same kind of ordering rule, and mine is short: Commit before you consult.
Write your own assessment and plan first. Then open the AI. Then account for every difference between the two.
That sounds like common sense. The research suggests skipping it carries a real cost, and that the order of operations protects judgment better than any training course I’ve seen offered to clinicians so far.
The anchoring problem
An AI recommendation that arrives before your own impression becomes the anchor. A Nature Medicine study tested nearly 800 participants, from laypeople to primary care physicians, with an AI dermatology tool that explained its reasoning. People who saw the AI recommendation first grew more deferential to it. Experienced physicians resisted bad advice far better than laypeople, yet the pull showed up even among doctors.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Radiology documented the same effect earlier. In a 2023 study of 27 mammogram readers, those with an average of more than 15 years of experience saw accuracy fall from 82 percent to 45.5 percent when the AI suggested the wrong category. Expertise softened the effect without removing it.
Education alone falls short
The usual answer is education, on the theory that clinicians who understand how large language models work will catch the errors themselves. A randomized trial put that to the test. Forty-four physicians, all formally trained in AI literacy, worked diagnostic cases with ChatGPT available on demand. The group fed suggestions with deliberately planted errors scored 73.3 percent on diagnostic reasoning. The group fed clean suggestions scored 84.9 percent.
Knowing a tool can be wrong and acting on that knowledge under load are separate skills, and anyone who has carried a full board late in a long shift knows which one gives out first.
What disuse costs
The second reason to commit first is maintenance. A skill you stop practicing decays. In a multicenter study in The Lancet Gastroenterology & Hepatology, experienced endoscopists who had grown used to AI polyp detection found fewer adenomas once the AI was switched off: 22.4 percent, down from a baseline of 28.4 percent. The design was observational, and the authors say other factors may have contributed. The signal still deserves attention from every specialty adopting decision support.
Committing first keeps the skill in use. Each case becomes a repetition you perform yourself, with the AI serving as a second reader rather than a substitute.
Why trainees need the rule most
For residents, the stakes run higher, and a 2026 Nature Medicine perspective names the danger “never skilling”: trainees who lean on AI through their formative years and never build baseline reasoning at all. Its proposed fix is a sequence. Reason without AI first, then learn to calibrate trust in machine output, then integrate AI under supervision.
That staging already governs procedures, since no program hands an intern a central line kit and walks away. Diagnostic reasoning deserves the same progression. This rule is the simplest way I know to build it into a shift.
What it looks like on shift
Picture a hypothetical patient, sixty-four years old, with chest pain and a nondiagnostic first ECG. Under this rule, the resident presents a differential and a disposition plan before anyone opens the tool. Then the team runs the case through the AI, and if it raises aortic dissection when the resident didn’t, that gap becomes the teaching point for everyone standing at the bedside. If the resident caught something the AI missed, that lesson matters just as much.
Three habits make the rule routine:
- Record your impression before the AI’s output appears. Researchers studying automation bias have proposed decision support that gives advice on request rather than up front. If your system displays suggestions automatically, ask whether that setting can change.
- Reconcile out loud. For learners, the differences between their plan and the AI’s plan are the curriculum.
- Keep some repetitions unassisted. Where a quality metric exists, track your performance with the tool off, not only with it on.
The bottom line
AI will sit in every exam room I work in for the rest of my career, and I welcome it. Used in the right order, it catches what I miss. Used in the wrong order, it does my thinking for me while I believe I’m still doing it. The rule costs about a minute per patient. Commit, then consult.
Harvey Castro, known as DR GPT™, is an emergency physician, author, keynote speaker, and physician futurist. He is advisor to the CEO of Phantom Space. He writes and speaks about artificial intelligence in health care, digital health, space medicine, and the future of human-centered care. Explore his speaking topics and availability.
He is the author of Bing Copilot and Other LLM: Revolutionizing Healthcare With AI, Solving Infamous Cases with Artificial Intelligence, The AI-Driven Entrepreneur: Unlocking Entrepreneurial Success with Artificial Intelligence Strategies and Insights, ChatGPT and Healthcare: The Key To The New Future of Medicine, ChatGPT and Healthcare: Unlocking The Potential Of Patient Empowerment, Revolutionize Your Health and Fitness with ChatGPT’s Modern Weight Loss Hacks, Success Reinvention, and Apple Vision Healthcare Pioneers: A Community for Professionals & Patients.
He shares updates on LinkedIn, YouTube, Facebook, Instagram, and X. #DRGPT
