Clinical Judgment Will Decide Whether AI Helps or Harms Medicine
AI is moving deeper into diagnosis, and the central concern is no longer whether the software can generate answers quickly. The more urgent question is whether physicians will still know when those answers are wrong.
Roman Giverts, CEO and founder of Vumedi, argues that medicine risks a dangerous imbalance if enthusiasm for automation outruns the training needed to judge it. In his view, the stakes are not abstract or limited to workflow efficiency. They are measured in patient outcomes, because an incorrect algorithmic suggestion can still look polished, confident, and clinically plausible.
The problem is not simply that artificial intelligence may make mistakes. Every diagnostic tool can fail. The deeper concern is that clinicians who lean too heavily on AI may stop developing the instincts needed to notice when the machine has drifted off course. Once that skill atrophies, the error does not announce itself. It quietly enters the exam room, the chart, and the treatment plan.
Peer Learning Remains the Bedrock of Diagnostic Reasoning
Giverts places unusual weight on a part of medical education that is easy to overlook when technology dominates the conversation: watching experienced physicians think through real cases. That kind of peer-to-peer learning does more than transmit facts. It shows how a clinician weighs uncertainty, revises a differential diagnosis, and recognizes when a pattern does not fit the expected script.
Clinical judgment is built in those moments. A resident or young physician sees not only the right answer but the process of reaching it, including the hesitation, the correction, and the second look at a confusing finding. Those lessons matter because AI rarely teaches doubt well. It can deliver a result with speed and confidence, yet confidence is not the same thing as accuracy.
In practice, a doctor who has learned through direct observation is better equipped to challenge a machine when the recommendation conflicts with the bedside picture. Without that foundation, the algorithm can become a silent authority. The danger is not that physicians will become passive in a single dramatic moment. The danger is gradual. Each unchallenged output makes the next one easier to accept.
The Risk Is Quiet, Not Spectacular
AI failures in medicine do not always appear as obvious falsehoods. More often, they may take the form of a reasonable but incomplete interpretation, a missed clue, or a recommendation that fits a pattern while ignoring the patient in front of the clinician. A system can be useful and still be wrong in ways that are hard to catch if the user has not been trained to interrogate it.
That is why Giverts’ warning lands beyond technology policy and into medical professionalism. If future doctors are trained to consume machine output rather than actively assess it, the profession could lose one of its core safeguards: the ability to recognize when the answer seems tidy but the clinical story does not support it. The bedside consequences are obvious. A wrong impression about the diagnosis can alter testing, delay treatment, or send care down the wrong path before anyone notices the deviation.
The concern also reaches beyond individual clinicians. Health systems that adopt AI at scale may assume oversight is built into the platform itself. It is not. Governance, review, and human skepticism still depend on clinicians who know how to think independently of the screen.
Why Medical Education Has to Stay Human-Centered
The interview points toward a broader educational question: what should future physicians be trained to preserve while medicine incorporates more automation? Giverts’ answer is clear. Training should not focus only on teaching doctors how to use AI efficiently. It should also protect the habits that allow them to dispute it.
That means preserving clinical apprenticeship, real-time case discussion, and exposure to expert reasoning. It also means resisting the temptation to treat AI as a substitute for judgment simply because it is fast or numerically impressive. Medicine has always depended on the ability to synthesize incomplete information under pressure. AI can assist that process, but it cannot replace the moral and cognitive responsibility of deciding when a recommendation fails to match the patient’s presentation.
There is a practical reason this matters now. As AI becomes more common in diagnostic workflows, the average clinician will encounter it repeatedly, not rarely. Repetition can normalize deference. If younger doctors do not keep practicing the act of questioning, the habit of skepticism may fade just as machine-generated recommendations become more embedded in routine care.
The Future of AI in Medicine Depends on What Doctors Keep
Giverts argues that the future of medicine depends less on whether AI can outperform humans in isolated tasks and more on whether physicians retain the skill to confront it when necessary. That framing shifts the debate. The point is not whether AI belongs in medicine. It already does, and its role is likely to expand. The real issue is whether medical education will preserve enough human judgment to catch the moments when automation is persuasive but mistaken.
A physician who can identify a flawed algorithmic output is not resisting innovation. That physician is performing the oldest clinical duty in medicine: protecting the patient from bad information.
If the profession fails to preserve that ability, the consequence will not be dramatic at first. It will show up as a slow erosion of diagnostic independence, one accepted recommendation at a time. And once that habit is lost, an AI system does not need to be malicious to cause harm. It only needs to be wrong, while no one still knows how to say so.