A doctor’s appointment has always involved more than a conversation. Even before the stethoscope comes out, a clinician is reading the room: how someone walks in, whether they wince while sitting, the color of their skin, the rhythm of their breathing. Then comes the part that is much harder to digitize: pressing an abdomen, feeling a swollen lymph node, checking whether a joint is warm, testing muscle strength, or noticing the strange firmness of something that should not be there.

This is the uncomfortable but necessary question hanging over the new generation of medical AI: an AI doctor may be able to listen, look and ask increasingly smart questions. But how can it physically feel pain, tenderness, swelling, heat, stiffness or other symptoms that depend on touch?

For now, the honest answer is simple: it cannot – not in the way a clinician can with their own hands. And that is precisely why the future of AI healthcare is unlikely to be a chatbot replacing the physical exam. It is more likely to be a system that knows its limits, uses cameras and connected devices to gather better clues, guides patients through parts of an examination, and tells them when the screen is no longer enough.

Google’s newest AMIE research offers a useful snapshot of this transition. The company says its research medical AI can conduct real-time video consultations, interpret visual and audio cues, guide elements of a virtual physical exam and reason through possible diagnoses. In simulated consultations involving patient actors, evaluators rated it favorably on areas including history-taking, diagnostic accuracy, management and communication. But Google also makes the important caveat: AMIE remains a research system and needs substantially more work before responsible real-world deployment.

The body is not just data

The popular picture of an AI doctor is usually cinematic: a calm voice asks about a cough, scans a face, checks a few numbers from a smartwatch and produces an answer. Some of that is already plausible. Cameras can capture a rash, a limp, visible swelling or changes in breathing. Microphones can pick up a cough. Wearables can provide heart rate, rhythm alerts, blood oxygen readings in some devices, sleep patterns and activity trends.

But a physical exam is not merely a checklist of observations. It is an active process of testing the body. A clinician may press gently, then more deeply; compare one side of the body with the other; feel whether a lump moves under the skin; assess how a patient reacts before, during and after pressure. That response – a tightening of abdominal muscles, a sudden grimace, a flinch, a description of pain radiating elsewhere – can change what the doctor does next.

Pain itself is also not a single measurable substance that can be downloaded from a nerve ending. It is a personal experience, shaped by injury, inflammation, the nervous system, stress, past experiences and context. A sensor may eventually measure pressure, temperature, movement or electrical activity. It still cannot directly access the subjective feeling of, “This is a dull ache,” or, “That sharp pain appears only when you twist.”

That gap matters because touch can reveal information that video cannot. Deep abdominal tenderness, guarding, rebound pain, organ enlargement, subtle temperature differences and the texture of a mass are difficult or impossible to verify through an ordinary phone or laptop call. Reviews of virtual physical examinations note that the evidence base is still developing, with substantial limitations in the available research, particularly for examinations conducted in homes or via mobile devices.

What AI can do today?

That does not make video-based medical AI useless. It just defines its role more realistically.

A well-designed AI system can take a more structured history than many rushed appointments allow. It can ask when the symptom began, what makes it better or worse, whether pain travels, whether there is fever, nausea, weakness, shortness of breath or recent injury. It can prompt a patient to show a rash in better lighting, walk across the room, raise both arms, check a pulse or press lightly on a specific area and describe the result.

In AMIE’s case, Google describes a system that can observe audio-visual signals and guide a virtual exam during a live consultation. That is not the same as claiming it can palpate an abdomen through a webcam. It means the AI can use sight, sound, questioning and patient-assisted maneuvers to decide what information is available remotely – and, crucially, what is not.

This is where AI may turn out to be most useful: not as a magical remote replacement for touch, but as a tireless clinical co-pilot. It can document symptoms, translate medical language into plain English, flag warning signs, help clinicians decide whether an in-person visit is urgent, and make a remote consultation more systematic than an unstructured video chat.

The key distinction is between observing and examining. AI is getting better at the first. The second still has stubborn physical constraints.

The workaround: turn the patient into part of the exam

Telehealth clinicians have already developed workarounds. They ask patients to point to pain, compare swelling on both sides, check capillary refill by pressing a fingernail, perform range-of-motion tests, monitor their temperature and pulse, or use someone nearby to help with a camera angle. In some cases, doctors ask patients to press on an area themselves and report whether it hurts, or to perform movement-based tests that can reveal whether a symptom worsens under strain.

That can be clinically useful, especially for follow-ups, medication discussions, uncomplicated skin concerns, certain respiratory complaints, mental healthcare and routine monitoring. But it is not a literal replacement for palpation. A patient does not have a clinician’s trained reference point for what “rigid,” “mobile,” “boggy,” “warm,” “pitting” or “concerning tenderness” feels like. And a person who is hurting may naturally press too lightly, too hard or in the wrong location.

The safest version of AI telehealth will therefore need to be comfortable saying: “I cannot assess that adequately from here.” That is not a failure of the technology. It is good clinical triage.

Machines that can touch

The longer-term answer may involve hardware, not just more capable language models.

Researchers are developing haptic systems that capture touch-related information: force, pressure distribution, vibration, texture and tissue stiffness. In medical robotics, this is often described as palpation sensing. A robot instrument or sensor presses on tissue, measures the response, and feeds information back to a surgeon or software system. Recent research reviews describe tactile and force sensors that can characterize tissue properties, though they also stress that clinical integration, standardization and validation remain difficult.

Imagine a future examination room in a pharmacy, clinic or home care setting. A patient places a handheld scanner on a swollen knee. The device measures surface temperature, circumference and movement. A connected stethoscope captures heart and lung sounds. A cuff records blood pressure. A camera tracks gait. A clinician – perhaps supported by an AI system – sees the combined information remotely.

That future is more believable than a humanoid AI doctor making house calls anytime soon. The breakthrough will not be one device that “feels pain.” It will be a network of tools that converts physical signals into reliable, interpretable data.

There are early demonstrations of remote haptic medicine, including research setups where a clinician interacts with a haptic device and a robotic system reproduces aspects of remote examination. But demonstrations are very different from deploying a safe, affordable and widely usable product in real homes.

Why “feeling” is the hard part

The engineering problem is bigger than putting pressure sensors on a robot hand.

First, touch is interactive. A doctor changes pressure, angle and location constantly based on what they feel and what the patient says. The system must respond in real time without causing harm.

Second, human bodies vary enormously. A sensation that is normal in one patient may be alarming in another. Body size, age, muscle tone, surgical history, skin thickness, underlying conditions and pain sensitivity all affect what a clinician finds.

Third, clinical touch is not only mechanical. A doctor uses context. A firm area in the abdomen might be less concerning if it is clearly a contracted muscle and more concerning if paired with fever, weight loss, vomiting or a specific pattern of pain. This is where AI could help: connecting sensor readings with history, images, labs and medical records. But that also raises higher stakes if the system gets its interpretation wrong.

Finally, there is the human side. A physical examination is intimate. Patients need to understand what is being measured, who can access the data, how long it is stored, and when a remote device should never be used instead of an in-person evaluation. A machine that touches a patient, directly or via a remotely controlled device, has to clear a much higher bar for safety, privacy, consent and trust than an app that merely asks questions.

The real test: knowing when to stop

The most responsible AI doctor may not be the one that sounds most confident. It may be the one that recognizes uncertainty early.

If someone has chest pain, new weakness on one side, severe abdominal pain, a rapidly spreading infection, major trauma, signs of internal bleeding or another urgent symptom, an AI should not try to improvise a clever remote substitute for hands-on care. It should recommend immediate in-person assessment or emergency help. Remote care cannot reliably perform every necessary exam, and telemedicine is not suitable for all clinical situations.

That is also the central lesson behind the current excitement around medical AI. Language models can make a health conversation more attentive, more accessible and potentially more useful. Video can add a surprising amount of clinical context. Connected devices can steadily reduce the information gap. But none of those things means software has suddenly gained human touch.

For the foreseeable future, AI will be able to ask, “Where does it hurt?” It may even learn to notice that you are uncomfortable before you say it. What it cannot yet do is place two trained fingers on the exact spot, feel the difference between tension and tenderness, and instantly decide whether that finding changes everything.

That final step still belongs to the physical world – and, usually, to a human clinician in the room.