How Should a Physician Respond When a Patient Arrives with ChatGPT?
In this article
A New Voice in the Exam Room: A Reassuring Response to Patients Who Bring AI Output
Today many patients come to their appointment not only with lab results but also with printouts of the questions they asked a chatbot such as ChatGPT and the answers they received. In the US, one in three adults turned to artificial intelligence (AI) for health information in the past year, and half of oncologists say their patients use these tools. This raises a question oncology clinics now face every day: when a patient says "the AI suggested a different treatment," how should the doctor respond? In this article, we explain what the evidence shows, the steps of a conversation that protects trust, and how patients can use these tools more safely. We covered the broader framework for AI use in medicine in our article on JAMA's rules on artificial intelligence.
- Short answer: Respond with curiosity, acknowledge the effort, read the output together and place it in the context of the patient's own situation. Starting with the tool's flaws or brushing the topic aside may lead the patient to hold back information next time.
- Why does it matter? Chatbots can be rated as more empathetic and readable than physicians' forum replies to cancer questions; however, about one third of their treatment recommendations contained elements not concordant with guidelines, and 12.5% contained fabricated recommendations.
- The real problem is context, not knowledge: In a randomized study published in 2026, the model on its own identified the correct candidate diagnosis 94.9% of the time, whereas people using the same model found it at most 34.5% of the time.
- For patients: Use the AI's answer as a list of questions to bring to your doctor, not as a treatment decision; do not share your identifying information, and contact your care team if you have urgent symptoms.
Why does this matter?
A cancer diagnosis is one of the moments in life when the need for information is most intense. The terms in a pathology report, staging, treatment options and side effects may not be fully understood in a short visit (terms in the pathology report, what do stage 4 and metastasis mean?). At midnight, a chatbot that patiently answers every question seems well placed to fill this gap. Indeed, in a US survey conducted in February–March 2026, 65% of AI users said they turned to these tools to get information quickly, and 41% to get an idea before deciding whether to see a doctor.
However, the same tools can also give wrong, incomplete or inapplicable information in a confident tone. In examples shared by a breast surgeon, one patient had come to believe that mastectomy would always mean longer survival, and another had received a reassuring answer from AI about a symptom that required urgent evaluation. Doctors cannot see these conversations in full; they cannot know what the patient asked or what details were left out. So the question is no longer "should patients use AI?" but "when they do, how do we put this information in the right context?" (common misconceptions about cancer).
How should doctors respond when patients bring ChatGPT?
The doctor should first ask with curiosity: "What did you ask, and what worries you most?" Then the doctor should acknowledge the effort to seek information, read the output together with the patient and sort the information into correct, wrong or incomplete, and not applicable to the patient. In the final step, the recommendations are placed in the context of the patient's stage, pathology and goals, and what the tool can and cannot do is discussed calmly.
Three different reactions, three different outcomes
When a patient shares a recommendation from AI, the doctor's first sentence shapes the rest of the conversation. Brushing the topic aside ("don't look things up online") or jumping straight into correction ("most of this is wrong") does not remove the patient's wish to seek information; it may only lead the patient to hide this information from the doctor. Yet the patient's wish to take part in their care is the most valuable ingredient of good shared decision-making (a holistic approach to the cancer patient).
Figure 1 · Same question, three different answers
Brushing it aside, correcting right away or responding with curiosity
The Society for Translational Oncology's 2026 guide recommends this approach in the same order: first ask about the patient's AI use, then acknowledge the effort, and set realistic expectations once trust is established. A similar voice came from Türkiye: Ahmet Dirican, a professor of medical oncology in İzmir, shared that he encourages patients to tell him when they have consulted AI, and that he discusses the information they bring by sorting it into correct, wrong or incomplete, and not applicable to the patient.
Figure 2 · What can a chatbot do, and what can't it do?
A good starting point, not the decision-maker
What does the evidence say?
The answers seem empathetic
In a study published in JAMA Oncology in 2024, physicians' answers to 200 cancer questions posted on a health forum (Reddit) were compared with the answers of three chatbots. The physicians who rated the answers found that the best chatbot scored higher than the physicians' forum replies for quality, empathy and readability. This result helps us understand why patients turn to these tools: the chatbot is not in a hurry, it notices the emotional side of the question and it uses plain language.
Physician or chatbot? Written answers to cancer questions
200 cancer questions from Reddit; 1–5 scale rated by physicians (higher is better). Best chatbot (colored) and physicians answering on the forum (gray).
Physician answers averaged 125 words; the best chatbot's answers ran to an average of 198 words. The scores were given by physicians, not patients, and the answers reflect short forum exchanges, not conversations in the exam room. Writing that seems more empathetic does not mean a more accurate medical decision.
However, the limits of this finding matter. The physician answers being compared were not conversations in the exam room but short messages written voluntarily on a forum, and the scores were given by physicians, not patients. A warmer tone does not mean a more accurate treatment recommendation.
Errors and fabrication in treatment recommendations
In a study published in 2023, also in JAMA Oncology, researchers put 104 questions about breast, prostate and lung cancer to ChatGPT 3.5 and compared the answers with the US NCCN treatment guidelines. Of the answers, 98% contained at least one guideline-concordant recommendation; however, about one third of them (34%) also contained recommendations not concordant with guidelines, and 12.5% of the answers included fabricated recommendations that did not appear in the guidelines at all. Even more striking, correct and incorrect recommendations were often intertwined within the same answer, in the same confident tone; even expert physicians fully agreed in their scoring on only 62% of the answers. This study was conducted with an older model, and models have improved since then; but the problem of "right and wrong in the same paragraph" remains relevant (evidence-based medicine).
The model knows, the user cannot find it
Perhaps the most instructive evidence comes from a preregistered randomized study published in Nature Medicine in February 2026. In the UK, 1,298 people were given 10 medical scenarios prepared by physicians; participants tried to work out what the condition might be and where they should seek care, using one of three large language models (GPT-4o, Llama 3, Command R+) or a method of their own choice. When the models read the scenarios directly, they identified the correct candidate diagnosis 94.9% of the time. People using the same models, however, found it at most 34.5% of the time; moreover, they did worse than the control group, who searched using their own methods.
The model knows, the user cannot find it: loss in passing on information
Nature Medicine 2026; 1,298 people, 10 medical scenarios, randomized study; rate of correct answers, %.
People using the model found the correct candidate diagnosis less often than the control group, who searched for information with the methods they would normally use at home (odds ratio 1.76, in favor of the control group). There was no significant difference between groups in choosing where to seek care. Models that scored highly on tests did not show the same success when talking with real people.
The researchers also examined the reasons for this loss: people described their symptoms incompletely in their first messages, the model did not always ask for missing information, it gave inconsistent answers to similar questions, and people could not tell the correct recommendation offered by the model apart from the other options. In short, the problem lies not only in the model's knowledge but in the conversation between the person and the model. This also explains why the doctor's place in the exam room cannot easily be filled: a good doctor asks for missing information, picks out what matters and fits the result into the patient's life (why does a second opinion matter?).
Figure 3 · Why does correct information get lost along the way?
From the model's knowledge to the person's decision
Why do patients ask AI?
According to a KFF survey published in March 2026, 32% of adults in the US turned to an AI chatbot for health information in the past year, on par with the share who use social media for health information. In a survey of 1,000 US oncologists conducted in August 2025, 50% of oncologists reported that their patients used these tools at some point during their treatment, 38% said this caused confusion and mistrust in patients, and 10% said it helped patients understand their options and prepare questions.
Why do people ask AI their health questions?
KFF survey, US, February–March 2026; people who used AI for health information in the past year, %.
In the same survey, 41% of those who used AI for health information uploaded personal medical information such as test results or doctor's notes; 77% of adults are concerned about the privacy of this information. Of those who asked a physical health question, 42% did not see a health professional afterward.
A conversation that protects trust: five steps
The steps below were prepared by combining the recommendations of international guides with our everyday clinical experience. The aim is to put the information in the right context without dampening the patient's wish to seek information.
Figure 4 · Responding to a patient who brings AI output
Curiosity, acknowledgment, reading together, context and next step
1. Ask with curiosity. A question such as "What did you ask the AI? What worries you most?" brings out the real concern behind the output. Often the patient's question is not "why wasn't I given this drug?" but "am I getting good enough treatment?" Some doctors now ask this routinely: "Have you asked an AI about this?" When asked without judgment, patients share more openly.
2. Acknowledge the effort. Saying "It's really valuable that you took the time to understand your illness" does not mean confirming that the information is correct; it means recognizing the patient's effort and worry. This small sentence makes it easier for later corrections to be accepted without turning into defensiveness (cancer-related distress and anxiety).
3. Read together and sort into three buckets. The information patients bring usually falls into three groups. Reinforcing what is correct with evidence builds trust; what is wrong or incomplete needs to be corrected with a source; and information that is correct but does not apply to the patient needs to be discussed by explaining why. For example, a recommendation that applies to metastatic disease may not be appropriate at an early stage, or a drug that requires a specific genetic alteration will not work in a tumor that does not carry that alteration (next-generation sequencing, how to read a genetic test report).
Figure 5 · The three-bucket method
Correct, wrong or incomplete, not applicable to you
4. Place it in the patient's situation. A chatbot does not have the full picture of the patient's pathology report, imaging, kidney and liver function, other medications and expectations from life. Saying "This recommendation applies in some situations; let's look together at how it fits your diagnosis and your goals" both takes the information seriously and shows why the decision is individual. Explaining that treatment decisions are often made by multidisciplinary tumor boards also strengthens this context.
5. Discuss the next step and the right way to use the tool. Once trust is established, the patient is more open to hearing what the tool can and cannot do. At this stage, it helps to talk about ways to ask better questions, how to protect personal information and what to do in case of urgent symptoms. Asking the patient to bring the output again to the next visit keeps the conversation open.
Example phrases
| When the patient says | The doctor might respond |
|---|---|
| "ChatGPT says this treatment isn't the best one." | "That recommendation may be right for some patients. Let's look together at how it fits your diagnosis, your stage and your goals." |
| "The AI suggested a different drug." | "AI can list options in general terms, but it doesn't know the details that make you different. Let me explain why this drug would or wouldn't work for your tumor." |
| "The AI said this symptom is nothing to worry about." | "I'm really glad you told me. Some symptoms need to be checked right away during treatment; let's take a look together now." |
| "The AI found this article." | "Let's look at the source together. AI sometimes cites articles that don't exist; if it's real, we'll talk about what it says." |
| "Should I trust AI?" | "It can be useful as a starting point, but we'll make decisions together based on your medical history, your test results and what's right for you." |
- Mocking or belittling: Remarks like "Where do you find these things?" embarrass the patient and push them to withhold information.
- Forbidding: Warning "Don't look things up online or on AI" doesn't work; the patient keeps looking and simply stops telling you.
- Getting defensive: Treating an AI recommendation as a challenge to the doctor's knowledge turns the conversation into an argument.
- Giving a long lecture on AI: Technical explanations given before the patient's real concern is addressed do not build trust.
- Agreeing without checking: Saying "Yes, that's right" is also a mistake; drug doses, supplements and sources in particular must always be checked.
The time problem and the doctor's own use of AI
Reading a long AI output line by line during a visit is not always possible. A practical solution is to ask the patient to highlight their three most important questions, or to suggest sharing the output with a nurse or patient navigator before the next visit.
Doctors themselves increasingly use AI for note-taking, writing patient letters or searching the literature. This makes transparency mutual: doctors should also be able to say openly how and for what purpose they use AI. Using approved AI tools with defined limits in diagnostic and treatment decisions is not the same as consulting a general-purpose chatbot (FDA-approved AI-assisted colonoscopy, AI that predicts response to immunotherapy, an AI prescribing pilot).
For patients: using AI more safely
Using AI chatbots is not wrong; they can help you understand your illness, learn the terms and prepare better questions for your doctor. What matters is to see the answer you get not as a treatment decision but as a starting point for a conversation with your doctor. How a question is asked also shapes the quality of the answer (what should you ask your doctor?).
Figure 6 · A better question, a more useful answer
A question with context that asks for sources, instead of a general one
Figure 7 · Safe use for patients
Before and after asking a chatbot
- Protect your personal data: In the US survey, 41% of AI users had uploaded medical information such as test results or doctor's notes. If you are going to upload something, first delete your name, ID number, address and medical record number, and check the tool's data use settings.
- Do not wait in an emergency: While you are receiving chemotherapy or immunotherapy, symptoms such as fever, shortness of breath, severe diarrhea or bleeding should not be put off because a chatbot answered "probably nothing serious."
- Be careful with supplement and alternative product suggestions: Chatbots may give information with weak evidence about herbal products and supplements; these products can interact with your treatment (do supplements affect treatment?, why are alternative products dangerous?, supplements taken without evidence).
- Evaluate clinical trial suggestions together: AI can list clinical trials that may be suitable for you; however, eligibility criteria are detailed and should be reviewed with your doctor (what is a clinical trial?, phase 1 trials).
- Notice the emotional impact: Reading survival statistics alone at midnight can increase anxiety. Average values do not describe your personal path; talk about these topics with your doctor and, if needed, with psycho-oncology support (depression and anxiety in cancer, fear of cancer coming back).
The situation in Türkiye
- There are no large-scale data showing how often cancer patients in Türkiye use AI chatbots. However, our clinical observations suggest that use is rising quickly, especially among younger patients and family members.
- Under Türkiye's Personal Data Protection Law (KVKK), health data are a special category of personal data. When uploading e-Nabız printouts or pathology reports to chatbots, it is important to delete identifying information and to be aware that the data may go to servers abroad.
- Most studies examining the accuracy of chatbots were conducted with questions in English. The accuracy of answers in Turkish has not been tested to the same extent; drug names, licensing and SGK reimbursement information may not apply to Türkiye.
- Whether a drug suggested by AI is licensed in Türkiye and what its reimbursement conditions are should be checked separately (new cancer approvals and AI in 2026).
Questions you can ask your doctor
- Which of the questions I asked the AI apply to my situation?
- Does the treatment recommendation in this output fit my stage and test results? If not, why not?
- Is this source real, and is it reliable?
- For which symptoms should I contact you or go to the emergency department right away, without asking AI?
- How could I bring my AI outputs to the next visit in a way that makes it easier for you?
- "AI knows more than the doctor." No. The model can access a wide range of information; however, it cannot examine you, cannot see your full medical record and may not pass on what it knows in tests with the same success in real conversations.
- "If it writes with empathy, it must be right." No. Warm, fluent language does not show that the information is correct or right for you.
- "My doctor will be offended if I say I asked AI." No. Telling your doctor is the safest way to have wrong information corrected and correct information reinforced.
- "If it cited a source, it must be right." No. Chatbots sometimes cite articles that do not exist or incorrect guideline information.
- "Using AI is entirely harmful." No. It can be useful for learning terms, preparing questions and rereading information; what matters is using it together with your doctor.
Frequently Asked Questions
- How should doctors respond when patients bring ChatGPT?
- The doctor should first ask with curiosity, acknowledge the patient's effort to seek information and read the output together with the patient. The information is sorted into correct, wrong or incomplete, and not applicable to the patient; the recommendations are placed in the context of the patient's stage, test results and goals. Brushing the topic aside or jumping straight into correction can damage trust.
- Is ChatGPT reliable for cancer treatment information?
- On its own, it is not a reliable source for treatment. In a 2023 study, about one third of ChatGPT 3.5's cancer treatment answers contained recommendations not concordant with guidelines, and 12.5% contained fabricated recommendations. It can be useful for understanding terms and preparing questions; decisions should be made with your doctor.
- Should I tell my doctor that I consulted AI?
- Yes. Telling your doctor allows wrong or inapplicable information to be corrected and correct information to be reinforced. Bringing the output in writing and marking your three most important questions makes the visit easier.
- Is it safe to upload my test results to a chatbot?
- It is not risk-free. Health data are a special category of personal data. Before uploading, delete your name, your T.C. national ID number and your medical record number, check the tool's data use settings and always discuss the interpretation with your doctor.
Glossary
- AI chatbot
- A computer program that answers written questions in human language; ChatGPT, Gemini and Claude are examples.
- Large language model
- An AI model trained on very large collections of text that generates text by predicting the next word.
- Hallucination (fabrication)
- When a model confidently produces information, a source or a recommendation that does not actually exist.
- Guideline concordance
- The extent to which a recommendation matches evidence-based treatment guidelines such as NCCN and ESMO.
- Shared decision-making
- The doctor and the patient reaching a decision by weighing the options, the evidence and the patient's values together.
- Randomized study
- A study in which participants are assigned to groups by chance, so that methods can be compared fairly.
- Context
- The personal details needed to interpret information correctly: stage, pathology, genetic testing, comorbidities and goals.
- Empathy score
- A rating of how well a response recognizes and takes into account the other person's feelings.
- Special category of personal data
- Under KVKK, health, genetic and similar personal information that is protected more strictly.
- Psycho-oncology
- The field of science and support that addresses the emotional and psychological needs of people with cancer and their loved ones.
DROZDOGAN Academy Commentary
There is one thing we learned in the "Dr. Google" era: a patient who seeks information is a patient who wants to take part in their care. With chatbots, the language of this wish has changed, but its essence has not.
⚡ Why it matters
A significant share of our patients now come to their visit having already had a preliminary conversation with an AI. An appointment that takes place without knowing the content of that conversation is an incomplete appointment.
✅ Strengths
When used well, AI helps patients learn the terms, organize their questions and come to the visit better prepared. Being patient and available around the clock can be valuable during anxious times.
⚠️ Limitations
The evidence shows that what models know can get lost in real conversations, and that correct and incorrect recommendations can be intertwined in the same answer. Data on the accuracy of answers in Turkish and on how often these tools are used in Türkiye are limited.
🩺 In practice
We ask our patients without judgment whether they have consulted AI, read the outputs together and sort the information into correct, wrong or incomplete, and not applicable to the patient. We are especially careful about urgent symptoms and supplement suggestions.
❓ Open questions
Will health-specific chatbots be safer for real patients? Do patients who prepare with AI make better decisions? How will visit times be adapted to this new need?
Bottom line: When a patient brings Dr. ChatGPT to the clinic, the best thing to do is not to close the door but to invite it to the table and read together. Putting information in context remains an indispensable part of medicine in the age of AI (both sides of the table, alternative medicine through a medical oncologist's eyes).
- This article was prepared in November 2026.
- The data are drawn from articles in JAMA Oncology (2023, 2024) and Nature Medicine (2026), the KFF survey from March 2026, an oncologist survey conducted in August 2025, and 2026 publications by ASCO, the Society for Translational Oncology and OncoDaily.
- The survey data are from the US; no large-scale data for Türkiye could be found. The model versions in the studies may differ from those in use today.
- The example phrases were adapted from recommendations in international guides and translated freely. The Academy commentary is our own.
- This article is not medical advice; make your health decisions together with your doctor.
References
- Chen S, Kann BH, Foote MB, et al. Use of artificial intelligence chatbots for cancer treatment information. JAMA Oncol 2023;9(10):1459-1462. doi:10.1001/jamaoncol.2023.2954. jamanetwork.com
- Chen D, Parsa R, Hope A, et al. Physician and artificial intelligence chatbot responses to cancer questions from social media. JAMA Oncol 2024. doi:10.1001/jamaoncol.2024.0836. healio.com
- Bean AM, Payne RE, Parsons G, et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nat Med 2026;32(2):609-615. doi:10.1038/s41591-025-04074-y. pmc.ncbi.nlm.nih.gov
- KFF. Poll: 1 in 3 adults are turning to AI chatbots for health information. March 25, 2026. kff.org
- eMarketer. Oncologists report about half of cancer patients are using AI for information (Impiricus and Klick Health survey, August 2025). emarketer.com
- ASCO AI in Oncology. STO guide offers scripted responses to common questions about chatbot recommendations. August 20, 2026. ascoai.org
- The ASCO Post. Upcoming ASCO AI in Oncology webinar turns chatbot printouts into better patient conversations. November 3, 2026. ascopost.com
- OncoDaily. Ahmet Dirican: When patients bring ChatGPT to the clinic, what should doctors do? October 8, 2026. oncodaily.com
- OncoDaily. Zacharoula Sidiropoulou: From Dr Google to Dr GPT – what has changed in my consultation room. October 18, 2026. oncodaily.com
- Medical Economics. 40 million people now use ChatGPT daily for health questions, OpenAI report finds (ChatGPT Health). January 7, 2026. medicaleconomics.com
Was this article helpful?
Your feedback helps us improve our content.
