Why eye care keeps a physician in the room
Ask will AI replace ophthalmologists and you have to split the job in two. One half is reading: retinal photographs, optical coherence tomography scans, visual field plots, charts and letters. Pattern software is genuinely good at that kind of input. The other half is surgery and examination. Cataract removal, retinal repair, laser treatment and injections are performed on a moving patient, with tissue measured in fractions of a millimeter and no second attempt.
Diagnosis in this job is rarely a single image. A scan suggests something; the physician checks it against the patient’s history, medications, other conditions and what the eye looks like under the slit lamp. Then comes the decision no model owns: whether to operate, when, and what to tell a patient who is frightened about losing sight. Consent, risk and the choice to wait are legal and human acts.
Coverage, our estimate of the share of task time AI can handle today, sits at 21 out of 100 for this job. The physical share of the work is high enough that our robotics tier for it is a dexterous humanoid, not a fixed machine on a bench. That hardware does not exist in clinics.
What software carries, what it assists, and what stays with people
The task list above sorts this job’s work into three groups, and the pattern is consistent. Tasks with a clean input, a checkable output and a tolerant error cost drift toward software. 0% of task time sits in the group AI can handle on its own.
A larger slice is shared work, where the model drafts or flags and the physician signs. That assisted group holds 33% of task time. Screening triage and documentation belong here: useful, time-saving, still reviewed. Of the exposed time in this job, our split puts 46.7% on the automation side and 53.3% on the assistance side, which is why the honest story here is erosion of specific tasks rather than a job disappearing.
The rest, 67% of task time, stays with the ophthalmologist: operating, examining, deciding and explaining. If you want the full definition of how that time is measured, read how coverage is scored.
What the evidence actually shows
Our quality parity grade for this job is D. That grade means there is no direct, published test of AI against practicing ophthalmologists across this job’s real task mix, so we publish no parity number for it. Diagnostic accuracy studies in retinal imaging exist, but a screening model scored on labeled images is not the same test as a physician managing a clinic list and an operating schedule.
What would settle it is narrow and specific: prospective trials that compare an AI system and a qualified ophthalmologist on the same patients, measured on outcomes rather than image labels, and covering surgical decision-making as well as detection. Until that exists, treat confident claims in either direction with care. Our full approach is set out in the scoring methodology.
The market data points the same way. The Bureau of Labor Statistics counts about 8,950 US ophthalmologists outside pediatrics, with median pay of $300,080 and projected employment growth of 4.5% between 2025 and 2035 (BLS, 2025). That is a small, aging-driven specialty with a long training pipeline, not a field being staffed down.
When the picture could change
Most likely after 2042 (8 in 10 of our scenarios). What that window measures is explained on the replacement-year method page.
Two things could pull it earlier. Autonomous screening already has a regulatory path in diabetic eye disease, and each approval that removes a physician from a routine loop moves real task time. Cost is the other lever: the AI side of our cost panel runs in the tens to low thousands of dollars, against a human cost range that starts above twenty thousand. Where a task is pure image triage, that gap is hard for a health system to ignore.
Two things hold it back. Surgical hardware is the first. Our robotics estimate puts 44.1% of this job’s work in the physical column at a dexterous humanoid level of capability, which is far beyond today’s assisted surgical platforms that a surgeon still drives. Liability is the second. Someone has to carry responsibility for a blinding complication, and in current US practice that someone is a licensed physician.
What to do: treat imaging AI as a second reader you audit, not as a colleague you trust by default, and keep a record of the cases where it was wrong.
How ophthalmologists stay needed
Lean into the parts of the work the task list leaves with people. First, operative skill and complication management, including the cases that go badly in the middle of a procedure. Second, the judgment call on whether to treat at all, where comorbidities, medication and patient preference outweigh any single scan. Third, the conversation: consent, prognosis and follow-up with a patient who may be losing vision.
Two skills raise your value alongside the tools. One is reading AI output critically, knowing where a model’s training population differs from your patients and where false negatives cluster. The other is clinic design, because the gain from screening software shows up in how a practice routes patients, not in the model itself.
Close work is worth watching too. See how the picture differs for optometrists, orthoptists and ophthalmic medical technologists, whose task mixes carry more measurement and testing. You can put any two of them side by side on the job comparison tool, or view the wider diagnosing and treating practitioners family and the healthcare sector page. For context on where this specialty sits among jobs that mostly need a person (our top band, Nah.), see the list of jobs least exposed to AI.