Imagine it's 1770. You are in Vienna. Everyone is talking about a machine that can think. It's an ornately carved wooden cabinet, with a life-sized figure sitting behind it in robes and a turban. It holds a chess piece in its hand. Over the following decades it travels the world and beats some of the smartest minds of the era (including Napoleon) and nobody can work out how.
It was a hoax. Hidden inside that cramped cabinet was a human chess master operating the machine from within. It wasn't a thinking machine. It was just a very clever, very uncomfortable man. But the hoax worked because people wanted it to be real.
Fast forward to 1997. The dream gets closer. IBM's Deep Blue sits across from Garry Kasparov, the best chess player alive. It wins. No trick this time, just raw computing power. Deep Blue searched millions of positions every second and picked the strongest option it could find.
The "brute force" tactic used by IBM was impressive, but it only works if the search space is small enough. It wasn't enough to work on Go, an ancient board game with more possible positions than there are atoms in the observable universe. No computer could calculate its way through that. To beat the world's best Go players, a machine couldn't just calculate. It had to learn.
So, Google's DeepMind team built AlphaGo. Instead of calculating, a machine studied patterns from millions of games humans had played, then played against itself repeatedly until it got better. In 2016 AlphaGo played Lee Sedol, one of the greatest Go players in history. In case you don't already know what happened, I won't tell you. I recommend you watch the documentary "AlphaGo". DeepMind released the full documentary for free on YouTube.
So, what does any of this have to do with your eyes?
Nobody wrote AlphaGo a rulebook for Go or told it which moves were good or bad. Instead, it learned from games and from playing itself. Medical AI works in a similar way. By studying huge numbers of fundus images, AI models have learned to spot diabetic retinopathy, detecting patterns in the images that a human eye can easily miss.
How Machines Learn to See
In the last post, we talked about Google's 2012 cat experiment. A machine was shown millions of unlabelled images and was never told what a cat looked like. Yet it learned to recognise patterns associated with cats. That kind of learning without labels or supervision is called unsupervised learning.
That's not how a diabetic retinopathy model learns, though. It learns in a similar way to a wine expert perfecting their palate over time. Rather than memorising rigid rules, it builds knowledge through experience. Instead of wine, it learns from retinal images. The task involves a team of eye care professionals labelling a huge number of fundus photographs, called training data. The machine looks at each image, makes a guess, and checks it against the label. Depending on whether it is correct, the pattern is reinforced or a small internal adjustment is made. Either way, it moves on to the next image. With every pass, it gets a little more accurate. This process is called training, and because a human is supervising the labels it learns from, it is called supervised learning. What comes out the other end is called a model. The model contains a set of parameters that captures patterns the machine has found useful for detecting diabetic retinopathy.
Next, the model is tested on images it has never seen before, and its performance is measured. In some studies, systems have achieved performance similar to that of an experienced eye care professional.
Up until now, we've been looking at patterns human experts already knew existed, like microaneurysms and exudates. But what happens when the machine starts spotting patterns no one ever labelled, simply because no human expert knew they were there in the first place? This opens up a whole new field, one we'll dive into next time.
When the Machine Met the Retina
Imagine a small regional town with one GP clinic and the nearest optometrist two hours away. A 58-year-old man with type 2 diabetes hasn't had his eyes checked in four years. A four-hour round trip and a day off work is the main barrier for him.
His GP clinic has an AI screening tool. A trained nurse takes two fundus photographs and, seconds later, the result comes back. Referable diabetic retinopathy. He is referred, treated, and keeps his vision. Without this tool, who knows what the future would have held for him?
In 2018, a retina-screening system called IDx-DR made medical history. It became the first FDA-authorised autonomous AI diagnostic system for a medical condition. It could detect more than mild diabetic retinopathy without requiring a clinician to interpret the images. No other field in medicine had pulled that off before.
That milestone was important because diabetic retinopathy is a leading cause of preventable blindness worldwide. The medical community already had the tools to treat it and save people's sight. The real bottleneck was getting patients screened early enough, and this technology addressed that hurdle. A nurse could run the test, with performance comparable to that of an ophthalmologist.
A randomised trial of young people with diabetes demonstrates the significance of this bottleneck. Half were offered an AI eye exam during their routine endocrinology appointment, and 100% of them completed this the same day. The other half were given a referral instead. Only 22% of this group had been screened within the next six months.
EyeArt took things further by adding clinical triage. The system doesn't just spot the disease, it also flags whether a patient can wait a few months or needs urgent care this week. In September 2025 the South-Eastern Norway Regional Health Authority, covering more than half the country, began rolling it out across its health system. It is among the most significant deployments of autonomous AI for eye disease detection. The aim was to increase the percentage of Norwegians with diabetes receiving a timely retinal exam from 55% to 95%.
There are many more examples of such tools being used at scale across the world. Google's ARDA system, short for Automated Retinal Disease Assessment, has now passed 600,000 screenings worldwide. In Singapore, a system called SELENA+ has been built into the national screening programme, which handles over 100,000 patients a year, to scale up what human graders can get through. Melbourne-developed Eyetelligence, now operating globally under Optain Health, checks for diabetic retinopathy, macular degeneration, and glaucoma from a single photo.
The hardware required for screening has been shrinking too. In 2024 the FDA cleared the Optomed Aurora AEYE. This was the first handheld camera able to run autonomous AI screening. A big advantage of this is not requiring an imaging room. In most cases dilation is not required and the result takes about a minute.
And the list keeps growing, with more systems cleared by regulators every year.
Think back to the man in that regional town. He was not even in the system before. Now, he needs ongoing monitoring, management, and follow-up, potentially for the rest of his life. Multiply this by everyone worldwide who's never had their eyes checked.
Technology such as this doesn't replace clinicians, it brings a whole new wave of patients into the system.
Are these tools perfect? Not even close. If a tool is trained mostly on lighter skin tones it may not perform as well in a clinic serving a diverse community. This is just one example of bias hiding in training data. How's it supposed to know something nobody ever taught it? It's important to stay critical, but more on that later.
Where this leaves us
Next time, we look at what happens when these systems start finding hidden health signals in an eye photo that we never knew existed.