If you say the words "artificial intelligence" to someone, what do they usually say? ChatGPT, probably. Maybe Gemini, or Claude if they've been paying closer attention. Or remember Skynet, from the Terminator?
A traffic light doesn't usually make that list. I want to start our journey of understanding AI somewhere much older and much simpler than ChatGPT.
Let's consider the traffic light. It senses when a car approaches and watches for a pedestrian pressing the button. It knows how long vehicles have been waiting.
Surely this isn't AI, you might be thinking. Well, we first need to look at the definition of artificial intelligence. There are several definitions floating around, and no single official one everyone agrees on. A working definition I like to use is this: artificial intelligence is any system that uses data to make decisions or take actions that would otherwise require human intelligence.
Now, getting back to the traffic light. It does all of these things. It uses data from pedestrians and vehicles, makes a decision to change lights and takes action for a period of time. In its simplest form, AI such as the traffic light is a set of instructions. In our world there are countless technologies that follow such logic, like an elevator or a vending machine.
AI isn't one technology. The field is complex and difficult to classify, but to begin with let's break it into three distinct groups. We'll start with the oldest, and still the most common.
Rule Based AI
A rule based AI is created when programmers write explicit instructions and the system follows them exactly. It doesn't learn or adapt. It simply and reliably executes a script. The traffic light we started with is one of these.
A fascinating example of rule based AI is a program called ELIZA. The program was built in 1966 by an MIT computer scientist named Joseph Weizenbaum. It was designed to mimic a psychotherapist, and the trick behind it was pretty simple. It scanned the input for keywords and returned a pre-written response, often turning the sentence into a question.
Weizenbaum published a real transcript in his original 1966 paper. In the dialogue, a young woman was typing into the terminal because her boyfriend had forced her to seek help, claiming she was depressed.
When she typed, "Well, my boyfriend made me come here," ELIZA had no understanding of what was happening, instead it had a rule sitting in a list somewhere. If the keyword "boyfriend" is entered, mirror the phrase back as a question. So that's exactly what it did. A little further on, she mentioned her boyfriend thought she was depressed much of the time. ELIZA triggered a second rule. The word "depressed" matched a line in Weizenbaum's script, and out came a pre-written reply, "I am sorry to hear you are depressed." It wasn't empathy. ELIZA was just following rigid instructions based on the input. The full exchange is worth a read if you have the time, see the sources below.
What surprised people wasn't the sophistication of the program, it was how convincing it could be. Weizenbaum walked into his office one day to find his own secretary deep in conversation with ELIZA, and she asked him to leave the room so she could speak with it privately. ELIZA had become something that people were pouring their hearts out to. Weizenbaum went on to become a critic of AI, unsettled by how easily people mistook it for the real thing.
There is a name for what was happening to that secretary, and to plenty of people since. Anthropomorphism, the psychological tendency to attribute human traits to non-human entities. In 1966, Joseph Weizenbaum's ELIZA triggered it with nothing more than a handful of simple rules. Today's sophisticated models trigger it even more powerfully, and it has consequences for how much a patient or a clinician ends up trusting a tool. More on that later.
Predictive AI
With predictive AI the system learns patterns from the data instead of following human written rules. The classic demonstration came in 2012, when Google researchers let an AI "watch" millions of screenshots from YouTube. Afterward, the researchers discovered the system had taught itself the concept of "cat", without human intervention. That shift, from being told the rules to learning the patterns itself, is the single most important idea in modern AI.
Generative AI
Generative AI is the newest of the three, and the one most people mean when they say "AI" today. Instead of just recognising patterns, these systems create new content from what they've absorbed during training. This includes text, images, music and even videos. An AI trained on Rembrandt's body of work can produce a painting in his style that never existed. ChatGPT can write you an essay. Neither of these systems is "thinking" in any way we'd recognise as human thought, but they are, by definition, demonstrating genuinely intelligent behaviour.
Rules. Patterns. Creation. Three groups of AI.
Two Kinds of Clever
Modern AI, especially the generative kind, is genuinely impressive. It can write, paint, hold a conversation and pass exams most people would struggle with. It's tempting to conclude it's simply intelligent, full stop.
But intelligence isn't one single thing. AI is remarkably good at processing vast amounts of data, recognising patterns and working tirelessly. Humans are better at empathising, reading the room and sensing what is wrong with a patient before they say a word.
That's the real case for augmented rather than artificial intelligence. AI extends what a clinician can do. It handles the data, we handle the care. It augments, not replaces.
All three groups are already in your clinic
None of this is new for eye care. All three groups are already sitting on your bench, some of them have been for decades.
Rule based AI has been your silent partner for years. Every time you run a threshold visual field, rule based AI is making decisions about gaze deviation, fixation losses and false positives. Every one of those is a rule written by a programmer, a script that must be followed by the machine in all circumstances. Your non-contact tonometer works the same way when it decides whether a reading is valid. So does your autorefractor when it confirms a refractive endpoint.
Predictive AI is where things get exciting. The same pattern recognition approach that found cats in YouTube videos is now used to detect diabetic retinopathy and a growing range of other eye diseases, with accuracy that matches or exceeds a highly trained clinician. Modern OCT software uses predictive AI to analyse retinal layers and compare scans over time to flag subtle progression.
There's another layer to predictive AI in eye care that deserves its own post. AI is now finding patterns in retinal images that reflect what is happening throughout the rest of the body. Patterns we never knew existed. The retina is the only place in the human body where blood vessels and nervous tissue can be seen directly and non-invasively. Imagine the possibilities. More on that soon.
Generative AI is probably the most immediately useful group for you right now. AI scribes can sit in with a consultation, listen, and produce a draft clinical note before the patient has left the chair. The same underlying technology can draft a referral letter in your voice, translate a summary into plain language, or produce it in another language entirely. What's underappreciated here isn't just the time saved, it's the cognitive load returned. Every referral letter drafted at the end of a long day is a small withdrawal from a finite mental reserve. Generative AI doesn't just give time back, it gives presence back. Done well, these tools don't replace the clinical relationship. They protect it.
Why the group matters
Why is it important to identify which group a tool belongs to? Because AI can fail, and the way that each group fails is different. Knowing how it fails helps you stay critical. That matters most with the last two groups, where failures are harder to spot.
Rule based AI fails obviously. If the situation isn't covered by a rule, it either does nothing or stops and flags the problem. Point an autorefractor at a dense cataract and it won't invent a prescription, it will return an error.
Predictive AI fails quietly. Its output is a confidence score, something like an 87 percent likelihood of referable disease. That score is only as good as the data the model learned from. A model trained mostly on one population can be confidently wrong on a patient who doesn't resemble that data, and nothing in the output will tell you it's struggling. A screening tool can return "no referable disease" with the same confidence whether it's certain or out of its depth.
Generative AI fails fluently. It can write you a clinical note that reads perfectly but includes a symptom the patient never mentioned. Nothing flags it. Everything it produces is a draft that needs checking.
Where this leaves us
So, what is AI? It's a traffic light that became a chatbot that became a cat detector that became the software reading a patient's retina and summarising a conversation. Three groups: rules, patterns, creation. And in eye care, all three are already open, and we will be seeing more of them.
We've only scratched at how these tools fail. Bias in training data, and why nobody can fully explain what an AI is looking at, are both subjects for future posts.
Next time you come across a new AI tool, ask yourself which group it belongs to. Rule based, predictive, or generative. Then ask yourself how you'd know if it got something wrong.
If you're after a place to start going deeper on AI, Professor Tim Rocktäschel's book, Artificial Intelligence: 10 Things You Should Know, is a great start. The audiobook is what got me hooked in the first place.