Why skin work stays with a doctor
From the outside, dermatology looks like an image problem: a photo goes in, a label comes out. That is a thin slice of the job. A skin check means looking at the whole body, feeling lesions, asking what has changed over six months, and weighing the drugs a patient already takes.
The procedures are the harder part. Shave and punch biopsies, excisions, cryotherapy, intralesional injections, laser work and Mohs surgery all happen with hands on a patient under local anesthetic. Each one needs consent, sterile technique, bleeding control, and a margin call made in the room while the tissue is open.
Then there is the treatment ladder. Starting a biologic or isotretinoin brings lab monitoring, program rules, prior authorization and a prescriber who carries the liability. A model can rank a differential. It cannot sign for the drug or own the outcome. The way we score jobs treats that responsibility as part of the work, not paperwork around it.
What AI does, what it assists, and what people keep
The work software can take on alone is mostly clerical and first-pass screening: drafting the visit note from the recorded conversation, handling coding and referral letters, and flagging lesions in a stored photo queue for a clinician to review. Share of task time in that group: 5%.
The assisted middle is bigger and more interesting. Image models give a second read on dermoscopy pictures, sort teledermatology submissions so urgent cases surface first, and summarize years of prior notes before a follow-up. A dermatologist still signs the diagnosis. Share of task time that sits in that group: 19%.
What stays with a person is the exam, the blade and the decision. Full-body inspection, biopsy and excision, post-op care, and the conversation about a suspicious mole on a worried patient’s back. Share of task time there: 76%. Our coverage figure, which asks how much of the task time AI can handle today, comes out at 20; how coverage is measured explains what counts.
What has actually been tested
Evidence grade for this job: D. There is no direct test of an AI system against working dermatologists across the whole job, so we publish no parity number for it. Benchmarks on curated lesion images test one task under set conditions. A clinic day is not that task.
What would settle it is a prospective study in real practices: unselected patients and skin tones, full-body checks rather than cropped photos, biopsy decisions and histology outcomes tracked, with AI alone, clinician alone, and clinician plus AI compared. Until something like that exists, assistive use is the measured ground and autonomous diagnosis is not. You can see how a parity grade is earned on the quality parity page, and what chatbots say about this kind of role in what the AIs say.
When the picture could change
Most likely after 2041 (8 in 10 of our scenarios). The replacement-year method sets out what that window does and does not claim.
Two things could pull it earlier. Regulatory clearance for tools that triage or read skin images without a clinician in the loop would change the workflow fast, because the software itself is cheap next to physician hours. Payer coverage for AI-read teledermatology would do the same, especially in places with long waits for a first appointment.
Two things hold it back. The procedural half of the job needs hardware that does not exist as a product: our robotics tier for this work is a dexterous humanoid, not a cart-based arm. And accountability sits with a licensed prescriber, so even a strong model needs a named doctor to approve the biopsy, the excision and the systemic drug. Demand is not falling either: the US Bureau of Labor Statistics counts about 11,370 dermatologists, with employment projected to grow 6.8% between 2025 and 2035, at a median wage near $328,730 (BLS, 2025).
How dermatologists stay needed
Lean into the parts of the job that cannot be sent as a photo. Procedural dermatology, including excisions and Mohs surgery, is the clearest example. Complex medical dermatology, where immune-mediated disease and systemic therapy interact with the rest of a patient’s health, is the second. The exam-room decision itself is the third: explaining risk, agreeing on a plan, and knowing when to biopsy anyway.
Two skills travel well. The first is reading model output critically, which means knowing where image tools fail, how confident scores behave on unfamiliar presentations, and when to overrule them. The second is running the workflow around them: triage rules for teledermatology, documentation standards, and who checks what before a result reaches a patient.
What to do: compare your own role with a neighboring one on the comparison tool before deciding what to learn next.
Nearby jobs face the same image-model question from different angles: radiologists, pathologists and family medicine physicians, who often make the first skin referral. For wider context, see the diagnosing and treating practitioners family and the healthcare sector page.
People asking will AI replace dermatologists are usually asking about a phone app that reads a mole. The honest answer is narrower and less dramatic: parts of the reading and most of the typing are moving, while the exam, the procedure and the signature are not. The headline Still needs a human figure for this job is 77 out of 100 (higher is safer).