Kerbside Consult

AI Has Entered the Consulting Room

The patient may still see the doctor. Increasingly, the consultation begins somewhere else.

Cyberdoc — writing on medicine since 1995

A photograph has been circulating on social media. It appears to show a notice at the entrance to a hospital:

“Patients who have already received a diagnosis based on ChatGPT are requested to seek a second opinion from Gemini, not from us.”

It is funny. It is also entirely believable.

There is just one problem: I cannot establish that the hospital — or even the notice — is genuine. The image may be staged, manipulated or AI-generated.

And perhaps that makes it the perfect way to begin an article about artificial intelligence in medicine.

In the age of AI, even an article asking whether we should trust artificial intelligence begins with a photograph we cannot entirely trust.

From Dr Google to Dr AI

The Internet changed medicine long before artificial intelligence arrived. For most of medical history, information flowed predominantly in one direction: from accumulated medical knowledge, through the doctor, to the patient.

The Internet disrupted that hierarchy. Patients could suddenly read about their illnesses before seeing their doctors. Medical journals, patient forums, pharmaceutical information and clinical guidelines became available to anyone with an Internet connection.

Doctors gradually became accustomed to the patient announcing: “I Googled my symptoms.”

Google, however, generally gave the patient information to search through. Generative AI does something rather different. It gives an answer.

And if the patient does not understand the answer, the conversation continues: “Explain that in simpler language.” “What does this mean for someone my age?” “Could my medicine be causing it?” “What questions should I ask my doctor?”

That is not merely a faster search engine. It is a conversational intermediary between medical knowledge and the patient.

The old sequence — medical knowledge → doctor → patient — became Internet → patient → doctor. Increasingly, it looks more like AI ↔ patient ↔ doctor.

There is a certain déjà vu to all this. In 1997, in Cybermed, I wrote about Information Online and Telemedicine; a year later, about Medical Search Engines. The question then was what would happen when medical information escaped the textbook, library and consulting room and became available to anyone with a modem.

Nearly thirty years later, the question has changed. The patient no longer merely searches for information. AI reads it, synthesises it, explains it — and sometimes arrives at the consultation having already suggested the diagnosis. The information revolution has become an interpretation revolution.

What patients are actually doing

Consider what patients now routinely ask general-purpose AI systems:

“My ferritin is low but my haemoglobin is normal. What does that mean?”
“My LDL cholesterol is high. Do I need treatment?”
“My child has developed a barking cough at night. What could it be?”
“Can these two medicines be taken together?”
“Explain this pathology report in ordinary English.”
“What questions should I ask my oncologist?”
“Should I go to the emergency department?”

Some patients upload laboratory reports, radiology reports, discharge summaries and photographs and ask AI to interpret them. In other words, what used to be the first clinical conversation may now be the second. The patient has already had a preliminary discussion with a machine.

The World Health Organization has explicitly identified patient-guided use — including investigating symptoms and treatment — as one of the major applications of large multimodal models in health. It also identifies diagnosis and clinical care, administrative work, education and scientific research as important uses.

What AI does surprisingly well

Doctors should resist the temptation to dismiss the technology simply because it unsettles the traditional hierarchy of medical knowledge. Some of its capabilities are impressive.

AI can explain medical terminology, summarise a discharge report, turn an impenetrable pathology report into understandable language, help a patient formulate questions before a consultation, and translate complex material into another language. It can do this instantly and repeatedly.

One much-discussed 2023 JAMA Internal Medicine study compared physician responses with ChatGPT responses to 195 patient questions posted online. Evaluators preferred the chatbot responses in 78.6% of assessments and rated them higher for both quality and empathy.

That study did not show that a chatbot is a better doctor. It evaluated written answers to online questions, not the messy reality of examining and treating a real patient. But it demonstrated something important: communication itself can be automated surprisingly well.

Much of the debate about AI in medicine asks whether machines will eventually diagnose disease better than doctors. That may not be the most useful first question.

One of AI’s immediate strengths is less glamorous but potentially transformative: translation.

Not merely translation between languages, but translation from medical language into human language.

Medicine has developed an extraordinary vocabulary: “subsegmental atelectatic change”, “indeterminate hypodense hepatic lesion”, “lymphovascular invasion”. To a clinician these phrases convey information. To a frightened patient they may convey little except anxiety.

A well-designed AI tool can potentially sit between the medical record and the patient and say: “This is what the report means. This is what it does not mean. These are the questions you may want to ask your doctor.”

The greatest immediate contribution of AI to patients may therefore not be replacing diagnosis. It may be translating medicine.

What happens when it is confidently wrong

Artificial intelligence can be wrong. Of course, doctors can be wrong too. Medicine has never possessed a monopoly on certainty.

But there is something particularly dangerous about the way generative AI can be wrong. It can be wrong beautifully.

The language remains polished. The explanation sounds authoritative. The paragraphs are perfectly structured. There may be no visible hesitation to tell a patient that the machine is uncertain.

The WHO has warned that large multimodal models can produce false, inaccurate, biased or incomplete statements. It also warns about automation bias: the tendency of patients and healthcare professionals to trust a machine’s recommendation and overlook errors they might otherwise have noticed.

The pattern of failure matters as much as the fact of it. A 2026 structured stress test published in Nature Medicine evaluated ChatGPT Health — OpenAI's consumer health tool, launched in January 2026 and already reaching millions of users — using 60 clinician-authored cases across 21 clinical domains. Performance was not uniformly poor. It was worst precisely where it matters most. The system correctly triaged classical, textbook emergencies such as stroke and anaphylaxis. But it undertriaged 52% of cases that physicians agreed required immediate emergency care: patients with diabetic ketoacidosis or impending respiratory failure were directed to "see a doctor within 24–48 hours" rather than the emergency department. The researchers also found that when a friend or family member minimised the patient's symptoms in the scenario, triage recommendations shifted significantly toward lower urgency — an odds ratio of 11.7 at the edges. The machine was not simply wrong at random. It was wrong in ways that mirrored the conditions most likely to mislead it: ambiguous early presentations, contradictory social context, deviations from the textbook pattern.

That leads to a deceptively simple question: How does a patient know when the machine does not know?

The missing patient

There is another limitation obvious to anyone who has practised clinical medicine.

The computer sees what you tell it. The doctor sees the patient.

A patient types, “I have some abdominal discomfort.” The doctor notices that he is pale and sweating. A parent says, “My child has a fever.” The paediatrician notices that the child is unusually lethargic. A patient describes “a little breathlessness”; the doctor notices that she pauses halfway through a sentence to breathe.

Medicine contains thousands of such signals: posture, expression, movement, skin colour, breathing, confusion, anxiety, the relative who keeps answering for the patient, the medicine that was forgotten, the symptom the patient is embarrassed to mention.

And sometimes there is that most difficult of clinical observations: something is not right.

Multimodal AI is increasingly capable of interpreting images, audio and other biological information. But clinical medicine remains more than processing the information deliberately supplied to a computer.

What the clinical trials tell us

A 2024 randomised clinical trial in JAMA Network Open produced a particularly interesting result. Fifty physicians worked through difficult clinical vignettes. Some had access to conventional resources; others also had GPT-4.

The physicians with AI access did not have significantly better diagnostic reasoning scores than those using conventional resources alone.

But in an exploratory comparison, GPT-4 by itself scored higher than the conventional-resources physician group on the study’s diagnostic reasoning measure.

That does not mean that doctors should be replaced by chatbots. These were simulated cases, not real patients. But it raises a more interesting possibility: perhaps the problem is not simply whether AI is good at medicine. Perhaps we have not yet learned how humans and AI should work together.

Giving a doctor access to a powerful tool does not automatically make the doctor better at using it. The stethoscope required training. Ultrasound requires training. Perhaps AI will too.

Doctors are using AI too

The picture of patients enthusiastically embracing AI while doctors remain sceptical observers is already outdated.

In a 2026 American Medical Association survey, 81% of physicians surveyed reported professional use of AI — more than double the 38% reported when the AMA first surveyed physicians in 2023. Uses included summarising medical research, creating patient instructions and documenting clinical encounters. That does not mean that 81% are asking a chatbot to diagnose patients. The category includes a much broader range of AI-enabled activities, from documentation and summarisation to patient instructions and other clinical workflows.

The enthusiasm is accompanied by caution. Physicians continue to express concerns about privacy, validation, liability and patients interpreting complicated results without medical assistance.

There is another question worth asking: what happens to clinical skill when a machine increasingly supplies the first differential diagnosis, drafts the first note and summarises the evidence? Calculators did not abolish mathematics, but we still expect students to understand arithmetic before they use one. Medical education may face a similar problem.

Could AI give us back the doctor?

There is, however, a more optimistic possibility.

Walk into many modern consulting rooms and watch where the doctor is looking. Too often it is not at the patient. It is at the computer.

Electronic medical records were intended to improve healthcare, yet documentation has become a major part of clinical work. Artificial intelligence may now help reverse some of that burden by drafting notes, summarising encounters and producing patient instructions.

Imagine a consultation in which the patient speaks, the doctor listens and the computer takes the notes.

The evidence that this is already happening is modest but real. A study published in JAMA in April 2026 tracked 8,581 ambulatory clinicians across five academic health systems — the largest controlled evaluation of its kind to date — and found that AI scribe adoption was associated with 16 fewer minutes of documentation time and 13.4 fewer minutes of total electronic health record time per eight hours of patient care. The gains were largest for primary care clinicians and for those who used the tools consistently. Those are not transformative numbers. But they point in the right direction: time reclaimed from the record and potentially returned to the patient.

For decades we worried that computers would make medicine less human. Artificial intelligence may, paradoxically, give doctors more time to behave like human beings.

That may prove to be one of its greatest contributions.

Privacy, accountability and the unanswered questions

People tell AI systems extraordinary things: symptoms, medicines, fears, sexual histories, mental health concerns and genetic information. They may upload laboratory reports, medical images and photographs.

A patient may disclose information to a chatbot that he has never told another human being. But does he know how that information is processed, retained, protected or transferred? The answers depend on the platform, its settings, the jurisdiction and the contractual arrangement.

A consumer AI service should therefore not automatically be assumed to provide the same confidentiality framework as a clinical relationship. The distinction matters: a consumer chatbot, a hospital’s approved and governed AI tool, and a regulated medical device are not the same proposition — legally, technically or ethically. A patient using a general-purpose AI at home operates without the validation requirements, audit trails or institutional accountability that attach to a clinical system procured and overseen by a healthcare institution. As AI becomes more deeply embedded in clinical workflows, that gap will need to be made explicit — to patients and to clinicians.

The challenge is straightforward to state and difficult to solve: how do we obtain the benefits of medical AI without turning the patient’s most intimate information into another digital commodity?

And what about Malaysia?

Malaysia is not a passive bystander to any of this. The Ministry of Health’s Malaysia Digital HEALTH Master Plan 2026–2055 explicitly envisages artificial intelligence, data-driven approaches, interoperability and digital governance as core features of the future health system — not as options to be considered later. Malaysia’s Personal Data Protection authorities have also developed guidance on automated decision-making and profiling, which has direct application to AI tools used in clinical contexts.

What this means in practice is that the same questions being asked in London, Boston and Singapore are questions for Kuala Lumpur too: who validates clinical AI before it is deployed, who is accountable when it errs, what governance framework distinguishes a consumer chatbot from a hospital-grade clinical tool, and how are patients informed about both the capabilities and the limitations of systems that may be advising on their care? A national digital health plan without clear answers to those questions is a plan for the technology, not for the patient.

Now imagine that an AI system recommends a diagnosis. The doctor accepts it. The diagnosis is wrong and the patient is harmed.

Who is responsible? The doctor? The hospital? The software company? The developer?

Medicine traditionally has a reasonably clear chain of accountability. The clinician makes the clinical decision and remains responsible for it. AI complicates that relationship because increasingly influential recommendations may emerge from systems whose internal reasoning is not transparent to either doctor or patient.

Responsibility cannot simply disappear into an algorithm. Someone must remain accountable to the patient.

Medicine is not simply an exercise in identifying the correct diagnosis.

A frail 88-year-old develops another cancer. Should we treat? A patient with advanced disease asks, “How long have I got?” A family says, “Don’t tell him.” A patient refuses a treatment the doctor believes could save her life. Another asks for treatment that is almost certainly futile.

No database solves these problems.

They require facts, certainly — but also judgement, values, experience, context and compassion.

A machine can generate words that sound empathetic. It can even generate words that readers rate as more empathetic than a hurried human response. But there remains a fundamental distinction.

The machine does not share responsibility for what happens next.

The doctor does.

Medicine is not merely the delivery of information. It is the sharing of uncertainty and responsibility between human beings. That is considerably harder to automate.

What patients and doctors should do

AI can be an excellent medical assistant for understanding terminology, summarising information, preparing questions and explaining general concepts.

A useful distinction is the difference between asking, “Help me understand this,” and asking, “Tell me what I should do.”

The closer the question moves towards diagnosis, prescribing, stopping treatment, interpreting a potentially serious symptom or deciding whether urgent care is required, the more important human clinical assessment becomes.

AI should not become a reason to postpone necessary medical care. Patients should also think carefully before placing identifiable or highly sensitive health information into consumer AI systems.

The mirror-image question deserves an equally direct answer: what should a doctor do when the patient walks in having already consulted an AI?

Neither irritation nor automatic acceptance is the right response. The sensible approach is probably this: ask to see what the AI said. Ask what information the patient gave it. Check whether the conclusion is reasonable. Correct what is wrong. Explain what the machine did not have — the examination findings, the clinical context, the things that changed the picture.

In that sense, AI literacy may become part of ordinary clinical literacy. Doctors once learned how to handle patients arriving with newspaper cuttings, then Internet printouts, then Google searches. The AI response on the patient’s telephone is simply the next thing brought into the room. The clinical task — taking it seriously, evaluating it and incorporating what is useful while correcting what is not — is recognisably the same.

The fourth person in the room

Every medical consultation has always involved three things: the patient, the doctor and the accumulated knowledge of medicine. Artificial intelligence has now given the third of these something resembling a voice.

It is extraordinarily knowledgeable. It is endlessly patient. It is available at any hour. And it will increasingly sit beside both patient and doctor.

That need not be something medicine fears. But neither should we confuse eloquence with accuracy, information with judgement, or simulated empathy with human responsibility.

AI has entered the consulting room. We are unlikely to persuade it to leave. Perhaps we should not even try.

The task now is considerably more difficult — and considerably more interesting:

We have to decide where it should sit.