Dermatologist — tasks, one by one

Examining the skin and judging lesions

Judging a single lesion is pattern recognition on an image, and image classifiers trained on dermoscopic photographs do that well. Whole-body photography systems can also map moles and compare them between visits. The examination itself is physical and selective, though. The patient has to undress, the light has to be right, a lesion may need to be felt, and the doctor has to notice the spot the patient never mentioned. History counts as well, such as a mole that has changed or cancer in the family. Licensing rules make the doctor answerable for a missed melanoma, so software comes in as a second opinion or a change-tracker that the dermatologist reads. It does not do the reading itself.

This says who owns the decision. It does not say how much of the dermatologist's time the task will take. Software that maps and pre-sorts lesions can shorten each examination and let fewer dermatologists see more patients. A dated deployment in which a classifier decides, with no dermatologist's review, which lesions in a specialist clinic are left alone would move this towards automating.

Triaging referrals and skin photographs

This is the most exposed part of the job because it is already done at a distance, from images and in bulk. The dermatologist sees a photograph and a few lines of text, not a patient. An image classifier can grade the same photograph in seconds. For lesions it rates as clearly benign, it can return an answer without a specialist ever looking. Health systems with long waits for skin cancer appointments have strong reasons to use it this way, and medical device rules can regulate software that makes such a decision on its own. The harder remainder stays with the dermatologist: unclear images, rashes, darker skin types where training data are thinner, and the cases the software passes up.

This says that the clearly benign end of triage can run without a specialist. It does not say how much of the work sits at that end. A missed cancer in an automated pathway could also bring back a rule requiring a human reader. The judgement would change with dated evidence that health systems withdrew autonomous triage after safety failures, or that its use stayed at pilot scale.

Biopsies and skin cancer surgery

This is fine manual work on a living body surface that moves and differs from person to person, and decisions about margins and closure are made during the cut. Surgical robots extend a surgeon's hands in large operations, but these are small procedures done under local anaesthetic in a clinic room. There a robot would cost far more than the procedure earns. Imaging software may help choose where to biopsy or help read margins. The cut and the responsibility for it stay with the doctor.

Keeping the work in human hands does not keep the volume. If triage and screening software sends fewer benign lesions for removal, the number of procedures could fall even though a person still does every one. Under local rules, part of this work can also pass to nurses and physician associates. A dated deployment of a device that performs skin excisions under supervision alone would move this judgement.

Diagnosing and treating skin disease

Inflammatory skin disease is diagnosed from a story more than from one picture: when it started, what makes it worse, what else the patient takes, how it responded last time. Image classifiers do worse on rashes than on moles because one appearance can have many causes. Language models can list possibilities, but they cannot examine a patient or answer for a prescription. Strong drugs need blood tests and side effects watched over time by a prescriber who holds the licence. Software helps at the edges, with reminders and the photographs patients send between visits, and does not change who decides.

This rests on how hard rashes are for current systems and on prescribing rules, and either can change. A general practitioner with decision support may also keep cases that used to be referred, which would shrink the dermatologist's share without any machine treating a patient. The judgement would move with evidence that a decision-support or image system diagnoses common rashes accurately across skin types and is used to close cases without referral.

Cosmetic and laser procedures

The machines here are tools the doctor aims. A laser sets its own energy, but someone has to judge the skin type, the risk of scarring or pigment change, and whether the patient's expectations are realistic. Paying clients choose the person and the clinic, and complications carry liability. Software that simulates outcomes or suggests settings helps with consultation and planning. It does not deliver the treatment.

The pressure on this work comes less from automation than from delegation. In many places nurses and aesthetic practitioners already do these procedures, and the rules on who may do them vary and change. This judgement does not cover that competition. A dated deployment of devices that deliver treatment with no operator beyond setup, in routine use across clinics, would move it.

Histories, notes, letters and patient advice

Speech recognition and language models now do listening, summarising and writing well. A tool that listens to a consultation can draft the note and the letter, and the same models can draft message replies and standard advice. Rules that make the doctor answerable for the record keep a person reviewing and signing, so the tools speed the work up without removing it. Advice that changes behaviour, such as persuading a patient to get a changing mole checked, still depends on trust in the person giving it.

Time saved on paperwork can go into seeing more patients instead of doing less work, so this does not say whether fewer dermatologists will be needed. The judgement would change with evidence that drafted notes are routinely filed without clinician review, or that dermatology clinics use AI drafting at scale with measured effects on time or errors.