The most useful AI answer may be the one that disagrees with the first AI answer. Medicine already works this way.
We seek second opinions.
Reviewers challenge authors.
Morbidity conferences examine decisions.
Clinicians ask what else a finding could represent.
Once a model produces a clear conclusion, both the model and the user can become anchored to it.
Interrupt the Anchor
The next prompts often deepen the original path: “Explain that more.” “Give me the evidence.” “Write the counseling.” Instead, interrupt it.
Ask: “Assume your conclusion is wrong. What is the strongest evidence-based case against it?” Then ask: “Which facts would discriminate between these two interpretations?”
This does not guarantee correctness. A model can generate weak arguments on both sides. But it can expose assumptions and alternatives omitted from the first response.
Use AI as a Red Team
After AI critiques a manuscript, reverse the role: “Defend the authors against each major criticism. Identify which criticisms remain valid after the strongest reasonable defense.” This matters because AI reviewers, like human reviewers, can overcall problems.
For a clinical policy, ask it to identify realistic ways the policy could cause unintended harm, inequity, delay, inappropriate escalation, or false reassurance. Then ask for measurable safeguards.
Challenge Answers You Agree With
Agreement is particularly dangerous.
If AI confirms what we already believe, the response feels intelligent because it resembles our reasoning.
That is exactly when to ask for the opposing case.
A useful instruction is: “I may be biased toward this conclusion. Do not reinforce my position automatically. Identify the strongest competing interpretation and the evidence that would favor it.”
Do Not Manufacture False Balance
Not every question has two equally defensible sides. Evidence may overwhelmingly support one conclusion. Ask for the strongest reasonable counterargument supported by evidence, and instruct the model to say so if no serious counterargument exists.
Clinical AI competence means structuring interaction so predictable human and machine weaknesses become easier to see. When AI gives you an excellent answer, do not immediately make it prettier. Make it attack the answer first.
TRY THIS: “Red-team your conclusion. Assume it may be wrong. Give the strongest evidence-based alternative, identify assumptions in the original reasoning, and list the facts or evidence that would discriminate between the positions. Do not create false balance if evidence clearly favors one side.”
A Three-Pass Method
I recommend a simple three-pass method for important AI-assisted reasoning.
Pass one is the ordinary analysis. Ask the model to assess the case, paper, policy, or argument using the information supplied.
Pass two is adversarial. Tell the model to assume its first analysis may be wrong. Ask it to identify the strongest competing explanation, the weakest assumption in its first answer, and the evidence that would most seriously challenge it.
Pass three is adjudication. Ask the model to compare the two analyses and state which disagreements can be resolved from the available evidence and which remain uncertain. Then verify the decisive claims yourself.
This is much more useful than repeatedly asking the same model, “Are you sure?” Models can respond to that question with another layer of confidence. A structured adversarial pass requires the system to produce a competing account.
The technique is especially valuable in peer review. AI may identify a supposed fatal flaw. On the second pass, ask it to construct the strongest defense the authors could reasonably make. If the criticism collapses under that defense, it was probably not a major criticism. If it survives, you have a stronger basis for raising it.
Human Bias Is Part of the Problem
The reason for red-teaming is not that AI alone is biased. Humans are biased too. We anchor. We seek confirming evidence. We become attached to diagnoses, theories, and manuscripts. Seniority does not eliminate these tendencies.
AI can amplify those biases when we use it as an agreement machine. A user who repeatedly asks for support for a preferred conclusion can obtain an increasingly persuasive argument for that conclusion.
But the same technology can be used in the opposite direction. It can be instructed to search for weaknesses, alternative explanations, and disconfirming evidence.
That is an important shift in mindset. The best use of AI may not be to make us more certain. It may be to make unjustified certainty harder to maintain.
A Practical Exercise
Try the red-team method on something you already believe strongly. That is where the exercise becomes uncomfortable and useful. Choose a clinical practice, research conclusion, or policy position you consider well supported. Give the model your argument and ask it to identify the strongest serious objection. Then require evidence for the objection and verify that evidence yourself.
Next, reverse the process. Ask the model to defend your original position against its own criticism. Finally, write your own adjudication without asking AI to choose the winner. Which claims survived? Which became more qualified? Which depended on evidence you had not considered?
This exercise teaches an important distinction between advocacy and analysis. Advocacy begins with a position and seeks the best argument for it. Analysis should remain open to the possibility that the position needs revision. Physicians do both in different settings, but we should know which mode we are in.
AI can make advocacy extraordinarily easy because it can generate persuasive support for almost any plausible position. Clinical AI competence means deliberately creating friction against that tendency. Sometimes the most valuable prompt is the one that makes it harder for us to keep believing what we already wanted to believe.
Closing Thought
The purpose of adversarial prompting is not endless skepticism. It is disciplined resistance to premature certainty. When an important conclusion survives a serious attempt to disprove it, our confidence is better earned.