Can an AI Be Your Doctor? What AI Does Well in Medicine, and Where a Clinician Is Irreplaceable

A friend texted a screenshot last week: a chatbot had, in her words, “figured out” her three-year-old’s rash in eleven seconds, complete with a confident name and a reassuring tone. The pediatrician, seeing the child in person the next morning, gently disagreed. The chatbot had never touched the skin, never seen the child scratch, never asked whether anyone at daycare was sick.

That small collision is why the phrase ai doctor is climbing the search charts as of September 2026. A wave of apps now advertise “free, anonymous AI doctors” with no sign-up, viral posts claim chatbots outdiagnose physicians, and hospitals quietly run AI on everything from mammograms to discharge summaries. The questions people type are blunt: Can it prescribe? Is it better than my doctor? Which one is best?

The honest answers are more interesting than either the hype or the backlash. AI genuinely does some medical work extraordinarily well. It also fails in ways that are easy to miss precisely because it sounds so sure of itself.

What does “AI doctor” actually mean?

Start with the words, because they are doing a lot of marketing work. Artificial intelligence, in plain terms, is software that finds patterns in large amounts of data and uses them to make predictions. Most tools sold as an “ai doctor” today are built on a large language model, a type of AI trained on enormous volumes of text to predict the most plausible next word, which lets it hold a conversation that reads like a clinician wrote it.

That fluency is the trick and the trap. The model is not reasoning from your body; it is generating the statistically likely continuation of the words you typed. When your words happen to match a textbook presentation, the output can be excellent. When your case is atypical, or you left out the detail that matters, the model will still produce something polished, because producing polished text is what it does.

Medical AI is broader than chatbots, though. A second family of tools, often called machine learning classifiers, looks at images or numbers rather than conversation. These read chest X-rays, flag suspicious spots on mammograms, or scan retinal photographs for diabetic eye disease. They tend to be narrow, trained on one job, and evaluated against a defined standard. A third family works behind the scenes: ambient scribes that listen to a visit and draft the note, or algorithms that watch hospital vital signs for early sepsis.

None of these is a doctor. A physician is a licensed person who takes a history, performs an examination, orders and interprets tests, weighs uncertainty, prescribes within a legal framework and, crucially, carries accountability for the result. Software has none of those properties, and no regulator in the United States, the United Kingdom or the World Health Organization’s member states has granted it any of them.

So when an app calls itself your AI doctor, read it the way you would read “gourmet” on a frozen dinner. It tells you about the marketing budget, not the ingredients.

What changed recently to make this trend now?

Three things converged. The first was scientific. In April 2023, a study in JAMA Internal Medicine compared chatbot answers with physician answers to 195 real patient questions posted on a public forum. Blinded clinician evaluators preferred the chatbot’s response 78.6% of the time and rated it more empathetic far more often. That paper, indexed on PubMed, became the citation behind nearly every “AI is kinder than your doctor” headline that followed.

In July 2023, researchers reported in Nature that a medical language model answered US licensing-exam-style questions at a level near the passing threshold and that clinician raters judged the large majority of its long-form answers to be in line with scientific consensus. Passing an exam is not practicing medicine, but the result shifted the conversation from “could it ever” to “what now.”

The second shift was regulatory and ethical. In May 2023, the World Health Organization issued a public call for caution, warning that large language models were being adopted faster than their safety could be evaluated. In January 2024, WHO followed with formal guidance on large multi-modal models in health, the models that handle text, images and sound together, setting out more than 40 recommendations for governments and developers and naming risks including false or biased outputs, privacy breaches and the erosion of clinical skills.

The third shift was commercial, and it is the one driving today’s searches. Consumer apps now advertise round-the-clock AI consultations, sometimes with a human clinician available in the background. Health systems have adopted ambient documentation tools widely enough that many patients have already been in a room where AI drafted the note.

The result, as of September 2026, is a gap between what the tools can demonstrably do and what the marketing implies. Closing that gap is the purpose of everything below.

What does AI genuinely do well in medicine?

Give credit where the evidence supports it. AI has earned a real place in several corners of medicine, and pretending otherwise is its own kind of myth.

Pattern recognition in images is the clearest win. Algorithms trained on hundreds of thousands of mammograms, retinal photographs or skin images can match specialist readers on defined tasks, and in screening programs they are being tested as a second reader that flags cases for a human. The machine never tires at the end of a shift, never skims the 400th film of the day. It also never notices the thing it was not trained to find, which is why the human stays in the loop.

Documentation is the second. Ambient scribes, which are AI systems that listen during a consultation and draft the clinical note, address one of medicine’s most corrosive problems: physicians spending as much time typing as talking. Early evaluations report that clinicians using them spend less after-hours time on records and feel more present with patients. Whether that translates into better outcomes is not yet proven, but reclaiming eye contact is not a small thing.

Prediction from routine data is the third. Hospital algorithms that monitor vital signs and lab values can raise an early alarm for deterioration or sepsis hours before a human might. Their record is mixed, with some producing more false alarms than useful ones, but the best-validated versions have been associated with earlier treatment.

Health literacy is the fourth, and the one most relevant to a person typing questions at midnight. A language model can translate a dense discharge letter into plain English, explain what a lab abbreviation means, or help you assemble the questions to ask at your next visit. That is a genuinely useful role, and it is very different from replacing the visit.

Notice the common thread. In every case the AI works alongside a clinician, doing a bounded task, with a human who owns the decision.

Can an AI doctor diagnose you the way a physician does?

To answer this, it helps to see how diagnosis actually happens, because it is not the process most people picture.

A clinician begins with a history, but the history is not just your words. It is your words plus the pauses, the thing you mention on the way out the door, the way you describe the pain with your hand rather than your vocabulary. Then comes the examination: the feel of an abdomen, the sound of a heart valve, the color of a fingernail bed, the reflex that is slower on one side. Roughly speaking, experienced physicians arrive at the working diagnosis from history and examination in the majority of cases, and use tests to confirm or exclude rather than to discover.

An AI chatbot has access to exactly one of those channels: the words you choose to type. It cannot press on your belly. It cannot notice that you are pale, or that your breathing is faster than your description suggests. It cannot ask the follow-up question that only makes sense after seeing your face fall when it mentions a particular word.

This is why studies that test AI on written case vignettes, where the whole story is already gathered and cleanly summarized, tend to flatter the machine. The hardest part of diagnosis, gathering the right information from a real person who does not know which details matter, has already been done by a human before the AI sees the text.

There is a second problem: the base rate. A chatbot answering a stranger has no idea whether it is talking with a healthy 25-year-old or a 70-year-old with heart failure and three medications. A physician who knows you carries that prior probability into every question. The same symptom, chest tightness, means something quite different in each of those people, and an algorithm that cannot weigh that difference is not diagnosing so much as pattern-matching on a fragment.

What the evidence actually says, graded by strength

Evidence in this field comes in three grades, and the grade matters as much as the headline.

Strongest: prospective and randomized studies of narrow tools. Some imaging algorithms have been evaluated in large screening populations and in randomized or controlled designs, with human readers as the comparison. Results are encouraging for cancer detection support and diabetic eye screening, with sensitivity comparable to specialists on the specific task. This is the most trustworthy evidence in medical AI, and it applies only to the tool tested, in the population tested, not to “AI” in general.

Middle: observational and benchmark studies of language models. The 2023 Nature paper reporting near-passing performance on licensing-style questions, and the JAMA Internal Medicine forum study in which evaluators preferred chatbot answers 78.6% of the time, are real and peer-reviewed. But they measure performance on text, judged by readers, without any patient outcome. The forum study compared chatbot answers with brief volunteer posts by physicians typing in their spare time, not with a consultation. Better-written is not the same as better care.

Weakest: vendor claims and viral anecdotes. An app’s statement that it is “as accurate as a doctor,” or a post describing one dramatic catch, is expert opinion at best and advertising at worst. The cases where a chatbot got it wrong are rarely screenshotted.

Two more findings deserve attention. Studies that gave physicians access to a language model during diagnostic reasoning have not consistently shown that the doctors got better; in at least one well-designed trial, the model alone scored higher on vignettes than the physicians, yet physicians given the model did not improve much, suggesting we have not yet learned how to use these tools well. And systematic reviews of medical chatbot accuracy repeatedly note that error rates vary widely by question type and that confident, fluent wrong answers occur in every model tested.

The fair summary: strong evidence for narrow, supervised tools; moderate evidence that chatbots write good answers; almost no outcome evidence that talking to one improves anyone’s health.

AI vs doctor accuracy: what each does better

People search “ai vs doctor accuracy” hoping for a single number. There is none, because accuracy depends entirely on the task. The table below summarizes where the evidence, and everyday clinical reality, place the advantage.

Task	AI strength	Clinician strength	Evidence grade
Flagging suspicious spots on screening images	Tireless, consistent, good sensitivity on trained targets	Judgment on unusual findings, decides what happens next	Strong (large prospective and controlled studies)
Answering written health questions	Fluent, thorough, often rated more empathetic in text	Knows your history, checks understanding, accountable	Moderate (blinded rater studies, no outcomes)
Diagnosing a real person with vague symptoms	Fast differential from typed description	Examination, context, follow-up questions, base rates	Weak for AI alone; core of medical training for clinicians
Drafting clinical notes	Saves documentation time, reduces after-hours work	Verifies accuracy, catches omissions	Moderate (early implementation studies)
Predicting deterioration from vital signs	Continuous monitoring, early alerts	Interprets alerts, acts on them, filters false alarms	Mixed (varies widely by system)
Prescribing and adjusting treatment	None; not legally permitted	Licensed responsibility, weighs risks for this patient	Not applicable; regulatory, not evidential
Delivering bad news, deciding goals of care	Can suggest wording	Presence, relationship, shared decision-making	Expert consensus

Two patterns jump out. AI’s advantages cluster around volume, consistency and language. The clinician’s advantages cluster around the body in the room, the story behind the symptom, and responsibility. Neither column replaces the other; the strongest results in the literature come from the combination, with a human deciding.

Is there a free AI doctor, and what does “free” cost?

Yes, in the sense that you can open a general chatbot or a health-branded app and type symptoms without paying. No, in the sense that a free tool does not become a doctor by being free, and the word “free” usually describes the payment, not the price.

The first cost is data. Anything you type into a consumer chatbot may be stored, reviewed to improve the model, or shared with partners, depending on the terms you did not read. Health information is among the most sensitive data you own, and most consumer chat tools are not covered by the medical privacy protections that bind your clinic. An app promising anonymity should tell you precisely what it keeps and for how long; if that page is hard to find, that is your answer.

The second cost is false reassurance. A polished paragraph saying your symptom is “most likely benign” feels like an answer. It is not an assessment. The model has no idea what it did not ask, and it has no way to notice that you are breathing hard while you type. Reassurance that arrives without an examination is the single most dangerous product a free ai doctor can hand you.

The third cost is the reverse: unnecessary alarm. Type a headache into a general model and you will likely receive a list that includes serious causes, because the model is completing a pattern in which serious causes appear. Anxiety, extra visits and unneeded tests follow.

Used well, a free tool is worth having. Ask it to explain a term, to help you organize your symptoms into a clear timeline, or to generate questions for your appointment. Do not ask it to tell you whether you are safe. That question requires a person, and preferably one who has met you.

Can an AI doctor prescribe medication?

No, and this is not a temporary limitation waiting on better software. It is a legal and ethical boundary that regulators have deliberately kept in human hands.

In the United States, prescribing authority belongs to licensed clinicians: physicians, and in defined circumstances nurse practitioners, physician assistants, dentists and others, under state law. In the United Kingdom, the same principle holds under the professions’ regulators. A prescription is not simply a recommendation; it is an act for which a named professional is accountable, one who has assessed the patient, considered allergies, kidney and liver function, interactions and pregnancy, and who can be contacted when something goes wrong. Software cannot hold a license, cannot be struck off and cannot be sued in the same way.

Some services blur this line by pairing an AI intake with a human clinician who signs the prescription. That arrangement is legitimate only if the human genuinely assesses you rather than rubber-stamping an algorithm’s output. Regulators in several countries have raised concerns about telehealth models where the “consultation” is a questionnaire and the clinician’s review lasts seconds. If an online service seems to produce prescriptions faster than a human could read your answers, be cautious.

A related question people ask is whether an AI can tell them to change or stop a medicine they already take. It should not, and you should not act on it if it does. Dose changes, tapers and stopping decisions involve risks that depend on your specific situation, and the only person positioned to weigh them is the clinician who prescribed the medicine or one who has your full record.

What AI can legitimately do is support the prescriber: flagging a potential interaction in the electronic record, summarizing a medication list, or drafting patient instructions for a clinician to check. In each case a licensed human reads, edits and owns the final decision, and that is exactly where the line should stay.

Where is a clinician irreplaceable?

Ask any physician what part of the job could never be automated and the answers converge on the same few things.

The examination. Hands on an abdomen can find a mass that no questionnaire would surface. A stethoscope can hear a murmur in a patient who came in for something else. Looking at how someone walks into the room, whether they wince sitting down, whether their skin has the faint yellow tinge that photographs never capture: these are data streams a chatbot does not have. Medicine has never stopped being a physical craft.

The unasked question. Patients routinely arrive with one complaint and leave having disclosed another, larger one, usually because a clinician noticed something and asked. Algorithms answer the question posed. Good doctors answer the question behind it.

Uncertainty held responsibly. Much of real medicine is not “this is the diagnosis” but “this is probably fine, here is what would change my mind, and here is when I want to see you again.” That is a plan built around a specific person’s risk tolerance, circumstances and ability to return. A language model does not know that you live alone, that you cannot afford a day off, or that your mother died of the thing you are afraid to name.

Accountability. When something goes wrong, there is a person whose name is on the chart, who answers the phone, who can be asked why. That is not a bureaucratic detail. It is what makes trust possible when the stakes are high.

And presence. The evidence that patients do better when they feel heard is not soft; it shows up in adherence, in fewer repeat visits, in how people cope with serious diagnoses. A chatbot can imitate warmth in text, and the 2023 forum study showed it does so convincingly. It cannot sit with you in silence after the scan result, and no one who has needed that would trade it for a well-composed paragraph.

Why does an AI doctor sound so confident when it's wrong?

The failure that worries clinicians most is not that AI makes mistakes. Humans make mistakes too. It is that AI makes them in a voice that does not waver.

The technical term is hallucination: a language model producing information that is fluent, specific and false, such as a study that was never published or a symptom association that does not exist. Hallucination happens because the model is optimizing for plausibility, not truth; a made-up citation with a realistic format is, to the model, a perfectly good continuation of the sentence. Every published evaluation of medical chatbots has found some rate of this, and the rate rises for rare conditions, recent developments and questions phrased with false premises.

Human experts signal uncertainty with hedges, pauses and “let me check.” Models trained to be helpful are, in effect, trained out of those signals. The result is what researchers call miscalibration: the model’s confidence does not track its accuracy. A wrong answer and a right answer arrive in identical tones.

Bias is the quieter cousin of this problem. Models learn from historical medical text and images, which underrepresent some populations and encode past inequities. Dermatology algorithms trained mostly on lighter skin have performed worse on darker skin. Language models have reproduced outdated race-based assumptions from older literature. WHO’s 2024 guidance names bias as a core risk precisely because it is invisible to the user; the answer looks the same whether the underlying data served your group well or badly.

Then there is automation bias, the human tendency to defer to a machine’s output even when our own judgment disagrees. Studies in radiology have shown that a wrong AI suggestion can pull an experienced reader toward the wrong call. If it can do that to a specialist, it can do it to a worried parent at midnight.

None of this means the tools are useless. It means their confidence is not information, and you should weigh their output the way you would weigh a very well-read acquaintance who never says “I don’t know.”

What happens to your health data when you use an AI doctor app?

The most under-asked question about any health chatbot is not “is it accurate” but “where do my words go.” The answer varies enormously, and the differences are not visible in the interface.

When your clinic uses an AI tool inside its own systems, whether a scribe drafting notes or an algorithm reading your scan, that tool operates under the same privacy rules that govern your medical record. In the United States that framework is HIPAA, the federal law protecting identifiable health information held by providers and their business partners; in the UK it is data protection law and NHS information governance. Vendors sign agreements, data use is restricted, and there is someone to complain to.

Consumer chatbots and many wellness apps sit outside that framework. If you download an app and type your symptoms, you are typically a user of a technology product, not a patient of a healthcare provider, and the protections are whatever the terms of service grant. Some tools state that conversations may be used to train future models. Some share usage data with advertising partners. Some claim anonymity while collecting device identifiers that make re-identification straightforward. WHO’s 2024 guidance flags exactly this: models deployed for health without health-grade governance.

A practical test takes about two minutes. Look for a plain-language statement of what is stored, whether humans review conversations, whether data trains the model, whether you can delete your history, and which country’s law applies. Absence of any one of these is meaningful.

Consider also what you type. A question about a medication is low risk. A full account of a mental health crisis, a sexual health concern or a genetic result is information you cannot take back once it leaves your device. If you would not want it read aloud by a stranger, do not hand it to software you have not vetted, however friendly its avatar.

How can you use AI health tools safely alongside your doctor?

Banning yourself from these tools is neither realistic nor necessary. Using them well is a skill, and it comes down to giving the machine the jobs it is good at and keeping the rest for people.

Use AI to prepare, not to decide. Before an appointment, ask a model to help you write a clear timeline of your symptoms, list your medications and supplements, and draft three questions. Clinicians consistently say that well-organized patients get better visits. That is a genuine improvement AI can deliver today.

Use it to understand, after the fact. Paste a de-identified sentence from a discharge letter or a lab report and ask what a term means in plain language. Then bring the explanation back to the person who ordered the test to confirm it applies to you.

Do not use it to triage an emergency. Chest pain, difficulty breathing, sudden weakness, severe bleeding, a child who is hard to rouse: these are phone calls to emergency services, not prompts. Every second spent typing is a second lost.

Do not use it to change treatment. If a chatbot suggests you adjust, add or stop a medicine, treat that as a question for your prescriber, never as an instruction.

Tell your clinician what you read. Many physicians now expect patients to arrive with AI-generated hypotheses and are glad to discuss them. Hiding it helps no one; a good clinician can explain why a suggestion does or does not fit you, and that conversation is often where the real learning happens.

Finally, notice how the tool handles uncertainty. A responsible health AI will say when it cannot know something, will point you toward a clinician for anything beyond general information, and will not pretend to examine you. If a tool never says “see a doctor,” that is not confidence. It is a design choice, and not one made in your interest.

Common myths about AI doctors, corrected

Viral claims travel faster than peer review. Here are the ones circulating most, and what the evidence supports.

“Chatbots are more empathetic than doctors.” The 2023 forum study did find blinded raters judged chatbot answers more empathetic than physicians’ written replies. But the physician replies were short volunteer posts on a public forum, not consultations, and empathy was rated in text by other clinicians, not experienced by patients. The finding says chatbots write warm paragraphs. It says nothing about being cared for.

“AI passed the medical boards, so it’s qualified.” Language models have performed well on exam-style questions. Exams test recall and reasoning on tidy vignettes. Licensure also requires supervised years with real patients, and the exam is the smallest part of what makes a doctor safe.

“AI will replace radiologists any day now.” This prediction has been made annually for a decade. What has actually happened is that radiologists use AI as a second reader while demand for their judgment has grown. Algorithms find what they were trained to find; a human decides what it means and what to do.

“An anonymous AI doctor is safer for privacy than a clinic.” Usually the reverse. Your clinic is bound by health privacy law. Many consumer apps are not.

“If the AI says it’s probably nothing, I can wait.” Reassurance without examination is the most hazardous output these tools produce. The model does not know what it did not ask.

“Doctors are against AI because they fear for their jobs.” Most physicians surveyed want better tools and already use them. Their concern is about unsupervised use by patients and about accountability, which is the same concern regulators have raised.

“There’s one best AI doctor.” No independent body ranks consumer health chatbots on patient outcomes, because no such outcome data exist. Claims of “most accurate” are marketing until a peer-reviewed comparison says otherwise.

Who checks whether medical AI is safe?

The honest answer is: it depends on what the AI claims to do, and the system was not built for tools that talk.

Software that makes a medical claim, such as detecting a condition on an image or predicting a clinical event, is regulated as a medical device in the United States, the United Kingdom and the European Union. Manufacturers must demonstrate safety and performance before marketing, and hundreds of such AI-enabled devices have been authorized over the past decade, the majority in radiology. This is real oversight, though it evaluates a defined function in a defined setting; a tool cleared to flag one finding is not thereby validated for anything else.

General-purpose chatbots occupy a different category. A model that answers any question, including health questions, has typically not been reviewed as a medical device because it does not claim to diagnose or treat, even when users treat it that way. Regulators on both sides of the Atlantic have acknowledged that this framework, built for single-purpose devices, fits poorly around software that does everything. Updating it is ongoing work, and as of September 2026 the rules remain in motion.

Internationally, WHO has set the ethical frame. Its 2021 guidance on AI for health laid out six principles: protecting autonomy, promoting human well-being and safety, ensuring transparency and explainability, fostering responsibility and accountability, ensuring inclusiveness and equity, and promoting AI that is responsive and sustainable. Its 2024 guidance on large multi-modal models translated those principles into specific recommendations, including that governments require independent evaluation before deployment and that developers involve patients and clinicians from the design stage.

Professional bodies add another layer. Medical associations increasingly publish guidance stating that clinicians remain responsible for decisions informed by AI, which means a doctor cannot blame the algorithm. That is exactly the accountability structure patients should want, and it is one more reason the human stays in charge.

When to see a doctor, whatever an AI told you

No chatbot output changes these thresholds. Some symptoms need a person, and some need one immediately.

Call emergency services now for chest pain or pressure, especially with sweating, nausea or pain spreading to the arm or jaw; sudden difficulty breathing; sudden weakness, numbness, facial drooping, confusion or trouble speaking; a severe headache that arrives like a thunderclap; uncontrolled bleeding; a seizure in someone without a known seizure disorder; a severe allergic reaction with swelling of the face or throat; signs of overdose or poisoning; a child or adult who is very hard to rouse; or any thought of harming yourself. Do not type these into an app first.

Seek same-day care for a fever with a stiff neck or a new rash, fever in an infant under three months, persistent vomiting that prevents you from keeping fluids down, severe abdominal pain, a wound that will not stop bleeding or shows spreading redness, sudden vision loss, or a medication side effect that is frightening or worsening.

Book an appointment for any symptom that has lasted more than two weeks without an explanation, unintended weight loss, a change in bowel or bladder habits, a mole that has changed, a lump anywhere, new or worsening mood changes, or a chronic condition that feels less controlled than usual.

Talk to your prescriber before acting on anything an AI says about starting, stopping or changing a medicine or supplement. This includes suggestions that sound reasonable. The risks of a change depend on your kidneys, your liver, your other medicines and your history, none of which a chatbot can see.

If you have already used an AI tool and it reassured you but something still feels wrong, trust the feeling. Clinicians describe this instinct as one of the most reliable signals they have, and it belongs to you, not to the software.

Frequently asked questions

Is there a free AI doctor?

There are free chatbots and apps that answer health questions, but none of them is a doctor in any legal or medical sense. A free tool can explain terms, organize your symptoms and help you prepare questions for a visit. It cannot examine you, know your history or take responsibility for advice. Check what the tool does with your data before typing anything sensitive, and treat its reassurance as a starting point for a conversation with a clinician, not a conclusion.

Can an AI doctor prescribe medication?

No. Prescribing is restricted to licensed clinicians under US state law and UK professional regulation, and no regulator has extended that authority to software. Some online services pair an AI questionnaire with a human who signs the prescription; that is legitimate only if the human genuinely assesses you. An AI also should not tell you to stop or change a medicine you already take. Any such question belongs to the clinician who prescribed it.

Is ChatGPT better than a doctor?

Not at being a doctor. General chatbots have performed well on exam-style questions and, in one 2023 study, wrote answers that clinician raters preferred to brief physician forum posts. Those results measure text quality, not care. A chatbot cannot examine you, weigh your history, order tests or take responsibility for what happens next, and it can state false information with complete confidence. It is a useful assistant for understanding and preparing, and a poor substitute for assessment.

Which is the best AI doctor?

There is no independently validated ranking, because no consumer health chatbot has published patient-outcome data that would allow one. Claims of being the most accurate are marketing until peer-reviewed comparisons exist. Instead of asking which is best, ask whether a tool is transparent about its data use, clearly states its limits, consistently directs you to a clinician for anything beyond general information, and has a genuine human clinician involved when it claims to offer care.

How accurate is AI vs a doctor?

It depends entirely on the task. Narrow imaging algorithms match specialists on the specific findings they were trained to detect, supported by large prospective studies. Language models score well on written vignettes where the history has already been gathered. On the hardest part of medicine, gathering the right information from a real person and weighing it against their history, there is no good evidence that AI alone performs well, and it has no way to examine anyone.

Is it safe to ask a chatbot about my symptoms?

It is reasonably safe to ask for explanations and to organize your thoughts, and unsafe to rely on it for a decision about whether you need care. Chatbots can both falsely reassure and needlessly alarm, and they answer only the question you typed. For emergency symptoms such as chest pain, sudden weakness, difficulty breathing or thoughts of self-harm, call emergency services rather than typing. Be mindful that what you enter into a consumer app may not be protected by medical privacy law.

Can AI read my X-ray or scan instead of a radiologist?

Not instead of, but alongside. AI algorithms authorized as medical devices can flag suspicious findings on images and are being tested as second readers in screening programs, with performance comparable to specialists on defined tasks. A radiologist still interprets the whole image, considers your clinical picture, notices findings the algorithm was not trained to detect and issues the report. The combination has generally outperformed either alone in studies.

Will AI replace doctors?

The evidence points to AI changing what doctors do rather than removing them. Tasks built on pattern recognition and documentation are being shared with software; examination, judgment under uncertainty, communication and accountability are not. WHO guidance explicitly frames AI as a support to clinicians and warns against eroding clinical skills. Predictions that specialists would be replaced have circulated for a decade without materializing, while demand for their judgment has grown.

What is an ambient AI scribe, and is my doctor using one?

An ambient scribe is AI software that listens to a consultation and drafts the clinical note for the clinician to review and sign. Many health systems now use them, and you should be told and asked for consent when one is active. Early studies suggest they cut after-hours documentation time and let clinicians make more eye contact. The clinician remains responsible for the accuracy of the final note, so you can ask to review what was recorded.

Should I share my medical records with an AI doctor app?

Be cautious. Records handed to your clinic’s own AI tools are covered by health privacy law; records uploaded to a consumer app usually are not, and may be stored, reviewed or used for training under the terms you accepted. Before sharing anything, find the app’s plain statement of what it keeps, whether humans read it, whether you can delete it and which country’s law applies. If any of those is missing, keep your records where the protections are.

References

• WHO: Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models (2024)
• WHO: WHO calls for safe and ethical AI for health (news release, 16 May 2023)

This article is for general information only and is not a substitute for professional medical advice. Please consult a qualified doctor about your individual situation.