Artificial intelligence has been working inside good eye hospitals for several years now, and most patients never notice it. It is not diagnosing anyone. It is reading images, spotting patterns across thousands of pixels, comparing today's scan against one from two years ago, and doing arithmetic that no human does faster or better. That is a genuinely useful set of jobs. It is also a much narrower set of jobs than the headlines suggest, and knowing the difference protects you from both over-trusting a report and dismissing a real warning.
The eye is unusually well suited to this
There is a practical reason ophthalmology moved first. The retina is the only place in the body where blood vessels and nerve tissue can be photographed directly, without cutting anything. A fundus camera produces a flat, standardised, high-resolution image. An OCT produces a cross-section of the retina accurate to a few microns. Feed a computer several hundred thousand of those images, each labelled by a specialist, and it becomes very good at recognising what they have in common.
That is all these systems do. They are pattern matchers trained on labelled pictures. They do not understand the eye. The distinction sounds academic until the day it matters.
Where the software genuinely earns its place
- Diabetic retinopathy screening. This is the clearest win. Algorithms reading retinal photographs can sort them into needs-a-specialist and does-not, with accuracy comparable to a trained grader. In a country where a large share of people with diabetes have never had their eyes photographed even once, a camera in a diabetes clinic that flags the worrying images is worth far more than a perfect system nobody can access.
- Change detection over time. Comparing this year's optic nerve scan to last year's, pixel by pixel, is tedious and error-prone for a human. Software does it consistently. Our OCT and Humphrey visual field analyser both run progression analysis of this kind, and it is one of the more reliable tools in glaucoma management, where the whole question is whether things are slowly getting worse.
- Lens power calculation before cataract surgery. Modern formulas use machine learning trained on large surgical outcome databases, fed by biometers such as the IOLMaster 700 and ARGOS. The result is that refractive outcomes after cataract surgery are noticeably more predictable than they were fifteen years ago. This is quiet, unglamorous AI and it has probably improved more lives than anything else on this list.
- Sorting and prioritising. Flagging which of forty scans taken this morning a doctor should look at first is a scheduling problem, and computers are good at those.
What it cannot do, and will not be doing soon
Every limitation below is a routine part of clinical work, not an exotic edge case.
- It cannot examine an eye. No algorithm measures your eye pressure, presses on your eyelid margins to see what comes out of the glands, watches how a two-year-old follows a toy, checks whether a pupil reacts sluggishly, or looks at the front of the eye through a slit lamp. A photograph of the retina is one page from a much longer file.
- It does not know your context. The same retinal appearance means different things in a 32-year-old with type 1 diabetes for twenty years, a pregnant woman whose retinopathy can accelerate quickly, and a 70-year-old with one functioning eye. Context changes urgency, and urgency changes what we do this week.
- It fails silently on unusual disease. A system trained mainly on diabetic and age-related changes will confidently call a rare inherited retinal dystrophy normal, because it has no category for it. It does not say I have not seen this before. It gives a number.
- It fails on bad images. A dense cataract, a small pupil, a watery eye or a patient who cannot hold still produces a blurred photograph. Some systems reject these. Others grade them anyway.
- It was trained on somebody else's population. Most published models were built on data from Europe, North America, China or Singapore. Indian eyes differ in pigmentation, disc size and disease mix. A model performing well in one setting can drift in another, which is why any system used on patients needs local validation rather than a certificate from elsewhere.
The failure mode that actually causes harm
These systems are wrong in a specific way: they are wrong with confidence. A doctor who is unsure says so, orders another test, or asks a colleague. An algorithm returns a clean percentage either way. There is no hesitation in the output, because hesitation was never part of what it learned.
This matters most in the fast-growing space of retinal photographs taken outside hospitals, at camps, in general clinics, at some optical shops. An automated report from one of these is a screening result, not a diagnosis. Normal on such a report means the software did not detect the patterns it was trained to detect in that one photograph. It does not mean your eyes are healthy, and it is not a substitute for a dilated examination. If you have diabetes, glaucoma in the family, high myopia, or you are over 40, you still need a proper comprehensive eye examination on the usual schedule.
What this means when you come in
Practically, very little changes for you. You will not be handed a diagnosis by a machine. What the software does is narrow the field so the specialist spends their time on judgement rather than on measurement, and so that a change of half a decibel on your visual field is not missed because two scans were printed six months apart.
If you are shown an automated report anywhere, three questions are worth asking:
- Was this a screening test or a full examination?
- Did a doctor look at the images themselves, or only at the summary?
- Was my eye dilated, and if not, how much of the retina was actually photographed?
The honest summary is that AI has made eye care more consistent, not more autonomous. It catches the boring mistakes that tired humans make at four in the afternoon, and it is useless for the decisions that need a person who has met you. Treat any claim beyond that with suspicion.
If it has been more than a year since your last full check, or you have diabetes and have never had your retina examined, that is the thing to act on. You can see how a full examination runs on our eye care page, or find your nearest centre on the branches page.
