An AI chatbot can help you prepare for a medical second opinion by explaining your diagnosis, listing alternatives worth asking about and suggesting questions. It cannot give a true second opinion, because it cannot examine you, review all your records reliably or take responsibility for its advice. Use AI to sharpen your questions, then get the second opinion from a qualified clinician.
In 2017 researchers at the Mayo Clinic looked at 286 patients who had been referred by primary care clinicians for a second opinion at the clinic’s general internal medicine division. They compared the referral diagnosis with the final diagnosis. Only 12% of patients had their original diagnosis confirmed as complete and correct. About two-thirds left with a refined or redefined diagnosis. And 21% left with a diagnosis that was distinctly different.
That study, published in the Journal of Evaluation in Clinical Practice, is from one specialist centre with a referred, and therefore more complex, group of patients. But it captures why second opinions matter. Diagnosis is hard, and a fresh pair of expert eyes often finds something.
Now a fresh pair of eyes is available to anyone with a phone, at no cost, at any hour. Patients are pasting their scan reports, pathology results and discharge letters into chatbots and asking: is this right?
The argument of this piece is that AI “second opinions” can genuinely empower patients, but they differ from a clinical second opinion in one crucial way. A human second opinion comes with accountability: a named clinician who has seen your records, can examine you and is responsible for their advice. An AI answer comes with none of that. The same tool can either sharpen your questions or leave you more anxious and less sure who to believe.
Are AI chatbots accurate enough to give a second opinion?
On written cases, the best models now perform impressively. On real patients, the evidence is thinner and more mixed.
A randomized trial published in JAMA Network Open in October 2024, led by Ethan Goh at Stanford, gave 50 physicians complex clinical cases to work through. Half had access to GPT-4 alongside their usual resources; half had conventional resources only. The physicians with the chatbot did not score significantly better on diagnostic reasoning: a median of 76% versus 74%. But the chatbot on its own, given the same cases, scored about 16 percentage points higher than the physicians using conventional resources.
That result is often quoted as “AI beats doctors.” It says something narrower and more interesting. On carefully written cases, with all relevant information laid out, the model reasoned well. And doctors given the model did not reliably benefit. The trial was not designed to explain why, but one widely discussed interpretation is that many physicians used the chatbot like a search engine, or discounted it when it disagreed with them. Having a strong second opinion available is not the same as using it well.
For patients, the gap is wider. A written case study contains exactly the information needed. A patient’s own records may be incomplete, contain jargon the patient cannot fully paste in, or leave out the crucial finding from the physical examination. Research on members of the public using chatbots for medical scenarios, including a 2026 Oxford study in Nature Medicine, found that people often failed to give the model the details it needed and struggled to tell good advice from bad.
What can an AI second opinion actually do well?
Used carefully, AI can make a patient much better prepared for a real second opinion. It is particularly useful for:
Explaining the diagnosis. What the condition is, what the terms in the report mean and how it is usually managed.
Listing alternatives. What other conditions can look similar, so the patient can ask: “Were these considered, and how were they ruled out?”
Identifying missing information. Which tests are commonly used to confirm the diagnosis and whether any seem absent from the records.
Outlining treatment options. The range of approaches commonly used, so the patient can ask why one was chosen over another.
Drafting questions. Turning a jumble of worries into a clear, prioritised list for the appointment.
Each of these supports a conversation with a clinician. None replaces it.
Where does an AI second opinion go wrong?
The failure modes are predictable.
False reassurance. A chatbot working from an incomplete summary may say the diagnosis seems reasonable when something important is missing from what it was shown.
False alarm. It may list rare but frightening possibilities without any sense of how likely they are for you, which can cause serious anxiety. Diagnostic lists without probabilities are a recipe for sleepless nights.
Confident errors. Language models can state things that are wrong with the same fluency as things that are right, and patients have no easy way to tell the difference.
No accountability. If a clinician gives a second opinion, they are professionally responsible for it, and there is a route to complain or seek redress. A chatbot’s disclaimers generally say it does not provide medical advice. If it is wrong, there is no one to hold accountable.
Damaged relationships. A patient who arrives saying “the AI says you’re wrong” can put a clinician on the defensive. A patient who arrives saying “I read that X is sometimes considered; was it ruled out?” usually gets a much better conversation.
A worked example: the thyroid nodule
This example is hypothetical.
James, 47, has an ultrasound showing a thyroid nodule. His doctor recommends monitoring with a repeat scan in a year rather than a biopsy. James is uneasy and pastes the ultrasound report into a chatbot, asking whether he should get a biopsy.
The chatbot explains how thyroid nodules are assessed, describes the scoring systems radiologists use to decide which nodules need a biopsy, and notes that most thyroid nodules are benign. It also says that, depending on size and features, some nodules are biopsied and others are monitored, and suggests he ask his doctor how his nodule was scored.
That is a useful response. It gives him vocabulary and a specific question.
But James also asks, “Could this be cancer?” The chatbot lists thyroid cancer types and survival rates. He spends the evening reading about rare aggressive forms. He sleeps badly.
At his follow-up, his doctor explains the nodule’s score, why it falls into the monitoring category under current guidelines, and what change would trigger a biopsy. James, now better informed, asks whether a referral to an endocrinologist for a formal second opinion would be reasonable. His doctor agrees. The endocrinologist confirms the plan.
The AI helped him ask the right question. It also caused a night of avoidable worry. The formal second opinion, from a clinician who had seen the images and was responsible for the advice, is what settled it.
The strongest counterargument: patients deserve every tool available
The case for AI second opinions is strong, and it comes from a real problem.
Diagnostic error is common across health systems. Formal second opinions are hard to get in many places, often expensive and slow, and more accessible to wealthy, well-connected and confident patients. Some families have described using chatbots to push for a diagnosis after years of being dismissed. For a patient who feels unheard, AI can provide the confidence and language to ask for more.
This argument deserves serious weight. The Mayo findings show that first diagnoses often change on review. Telling patients simply to trust their doctor ignores that evidence.
But the conclusion is that patients should use AI to ask better questions and to seek formal second opinions when there is doubt, not that AI can stand in for that second opinion. The accountability gap is not a technicality. It is the difference between advice that someone is responsible for and advice that nobody is.
What this means for patients, clinicians and health systems
For patients. Use AI to understand your diagnosis and prepare questions. If you still have doubts, ask your clinician for a referral for a second opinion; in many health systems you have a right to request one. Do not start, stop or change treatment based on a chatbot’s view, and seek urgent care if your symptoms worsen.
For clinicians. Expect patients to have consulted AI. Asking “what have you read, and what’s worrying you?” turns it into a shared starting point. Explaining your reasoning, including what you ruled out, often resolves more than any chatbot could. And if a patient raises a reasonable alternative, taking it seriously is good medicine. We have written about the risk of clinicians simply agreeing with AI drafts; the opposite risk, reflexively dismissing a patient’s AI-informed question, deserves equal attention.
For health systems. If patients are seeking AI second opinions because formal ones are hard to get, that is a signal about access. Clear, timely pathways to specialist review would do more for diagnostic accuracy than any consumer chatbot. And any AI tool offered to patients for diagnostic review should be tested on the narrow tasks where it is reliable, as we argued in AI diagnostics work best where the question is narrowest. An AI opinion that raises a worry with no pathway to resolve it is not empowerment. It is a new source of demand with nowhere to go.
Frequently asked questions
Can ChatGPT or another AI give me a second medical opinion?
AI can help you understand your diagnosis, identify alternatives worth asking about and prepare questions, but it cannot give a true second opinion. It cannot examine you, may not see your full records and is not accountable for its advice. For a real second opinion, ask your clinician for a referral to another qualified specialist.
How often do second opinions change a diagnosis?
In a 2017 Mayo Clinic study of 286 patients referred for a second opinion, only 12% had their original diagnosis confirmed as complete and correct, about two-thirds had it refined, and 21% received a distinctly different diagnosis. These were referred, often complex cases at one centre, so the numbers will not apply everywhere, but they show why second opinions can be valuable.
Is AI more accurate than doctors at diagnosis?
On some written case studies, the best AI models now perform as well as or better than physicians. A 2024 JAMA Network Open trial found GPT-4 alone outperformed doctors using conventional resources. But real patients are messier than written cases, and studies of the public using chatbots show much weaker results. AI is not a replacement for a clinician’s assessment.
Will my doctor be offended if I use AI to question their diagnosis?
Most clinicians are used to patients researching their conditions. How you raise it matters. Asking “I read that X can look similar; was it considered?” invites discussion. Saying “the AI says you’re wrong” tends to cause defensiveness. If you remain unsure, asking for a formal second opinion is reasonable and common.
How do I get a formal second opinion?
Ask your treating clinician for a referral, contact your health insurer or health service about second-opinion policies, or approach a specialist centre directly where that is allowed. Bring your records, imaging and test results, and a clear list of questions. AI can help you prepare that list, but the opinion itself should come from a qualified clinician.
The most useful thing an AI second opinion can produce is a better question for a human. The least useful is a new fear with no one responsible for answering it.
Have you used AI to check a diagnosis? Did it bring you closer to your clinician, or further away?
