AI in eye care: 27 of 29 results in a review top 90%. Who is liable?

AI in eye care does well in research. A systematic review published on 18 September in the International Journal of Ophthalmology read 34 studies of AI reading fundus photographs. In its own performance table, 27 of the 29 results are above 90%, though the column it labels accuracy also holds sensitivities and an area under the curve.

What it does not answer is who carries the risk. It names medicolegal liability as a concern, but its results give three barriers and no analysis of who is liable. Eight days earlier, an independent commission set up by the MHRA reported on exactly that, and the GOC says it will respond.

What the review did

A PRISMA review, not a meta-analysis. Scopus, PubMed and Web of Science were searched to June 2025, and 34 studies were included, three of them from the UK. Diabetic retinopathy was the commonest subject, in 25 of them. Studies using OCT or the slit lamp were excluded. On the Mixed Methods Appraisal Tool, 22 studies were moderate to high quality and 12 showed notable limitations.

What it found, and what it left out

Three barriers. The software does not fit the working day: separate programs, manual uploads, hardware that costs money. Clinicians cannot see how a black box reached its answer, which the authors tie to low trust. And models trained on one population do not travel, so performance falls in the groups already least well served. Their summary is blunt.

AI performs well in research (>90% accuracy) but falters in real-world use.

One hospital's simulation put numbers on the trade. A study published in PLOS Digital Health on 15 September let AI clear its own negatives on 6,904 fundus photographs read by three retina specialists. The images needing a specialist fell to 7.3%–8.0% for diabetic retinopathy, and mean sensitivity for it fell from 0.9168 to 0.8838. In an exploratory analysis across all three diseases, 27.7%–30.1% still needed one. A stricter threshold won back most of the diabetic retinopathy loss, so where to set it is a clinical judgement as much as a technical one.

Who is liable when AI in eye care is wrong

The National Commission into the Regulation of AI in Healthcare reported on 10 September with 44 recommendations. Under negligence as it stands, it found, claims are disproportionately likely to be brought against healthcare professionals and providers, because they have the clearest duty of care. It calls the result a perceived “liability sink”. Recommendation 28 is that the contract between manufacturer and provider should set out who delivers which risk controls. That is a recommendation, not a requirement.

Optometrists have worked this out. In the General Optical Council's 2026 registrant survey, 76% of the 2,064 optometrists who answered were concerned about their legal responsibility if something goes wrong, as many as were concerned about the AI making errors, and 65% rated their own knowledge of AI as poor.

Why it matters in practice

What changes is not the consulting room but what you ask a supplier and what you write down. The College of Optometrists' interim position says, among other things, three you can act on. Check the tool is registered with the MHRA as Artificial Intelligence as a Medical Device before you buy it. Tell patients when AI is being used, and where possible let them opt out. And keep records of what the AI provided, how it informed the decision, and when and why you preferred your own reasoning. That last one you can do this week.

The contract is the next renewal's job. The third is what to tell the patient who asks whether a computer will read their photographs. In NHS diabetic eye screening in England it will not yet, though the UK National Screening Committee is closer than it was: the evidence summary it consulted on concludes that some systems are as good as a first human grader and could be tested in real NHS services. That consultation closed on 3 August 2026, and no updated recommendation had been published by 1 October. For anything else, the answer is whatever your own practice's software does. What reaches the high street, to my mind, arrives inside a pathway, as Denmark's OCT roll-out did, not as a box on its own.

The caveats

• The review's arithmetic does not hold together: the abstract says 34 studies from 2018 to 2025, the results say 2016 to 2025, and the table lists 29 entries, nine of them dated 2026. One of those reads fundus and OCT together, which its own exclusion criteria rule out. It states no limitations of its own.
• The table's figures are each entry's own, mostly on curated data, never pooled, and not all are accuracies: some are sensitivities, one an area under the curve. Two rows cite reviews; the lowest, 73.0%, cites a news report. One model fell from 98.84% to 87.13% on outside data. None measures sight saved.
• The Commission is an independent expert advisory body set up by the MHRA. Its 44 recommendations are not law, the government response is still to come, and the liability passage is analysis, not law.
• The GOC figures are self-reported by the 3,451 who answered, an 11% response rate. The triage simulation is single-centre, retrospective and Korean, part-funded by the software's maker, two of whose employees are authors.

Sources

1. 1.Why aren't we using AI in eye clinics? A systematic review of barriers and solutions in AI-based fundus image diagnostics for ocular diseases · International Journal of Ophthalmology, 18 September 2026
2. 2.National Commission into the Regulation of AI in Healthcare: Recommendations for a future regulatory framework · GOV.UK, 10 September 2026
3. 3.Optical professionals cautiously optimistic about AI but raise concerns about errors and accountability, GOC survey finds · General Optical Council, 2 September 2026
4. 4.Interim position on the use of artificial intelligence (AI) in eye care · College of Optometrists, 9 June 2025
5. 5.Prospective evaluation of an artificial intelligence-enabled algorithm for automated diabetic retinopathy screening of 30 000 patients · British Journal of Ophthalmology, 30 June 2020
6. 6.UK NSC consults on use of artificial intelligence in the diabetic eye screening programme · UK National Screening Committee, 11 May 2026
7. 7.Evaluating an AI-assisted triage workflow for retinal diseases · PLOS Digital Health, 15 September 2026
8. 8.Registrant workforce and perceptions survey 2026: research report · General Optical Council, 2 September 2026
9. 9.Registrant survey 2026: profile of an optometrist · General Optical Council, 2 September 2026