What AI can and can't do: honest boundaries
A simple rule: if a person solves the task in a second and there are plenty of similar examples around, AI will probably handle it. If the situation is new and unlike anything it has seen, it gets it wrong — confidently. We find that line using self-driving cars and medicine. And work out how to decide up front whether to hand a task to AI at all.
From the last lesson you know AI's strength runs into its data. Now let's draw an honest map of the boundaries — what AI can really do, and what's beyond it in principle. That's the answer to "when not to use AI".
The main rule: the one-second test
Keep one simple rule in mind and it'll save you years of illusions: if a person solves the typical version of a task in about one second of thought — and the world is full of examples of it — AI will probably handle it. Spotting a car in a photo, hearing a word, telling spam from not-spam, recognising a face — these are all "one-second" tasks of perception, and there's an ocean of data for them. AI is strong here.
Now the flip side: predicting a stock price a month out from past prices is not a one-second task, and more importantly, past prices don't predict future ones. A double failure: the thinking takes a long time, and there's no pattern in the data anyway. The rule is crude but reliable: AI is a brilliant specialist in instant recognition, not a sage weighing up consequences.
Paired examples: can / can't
Let's calibrate on two fields where being wrong has a real cost.
- A self-driving car. Can: work out from camera and radar where the other cars on the road are — there are millions of labelled examples. Struggles enormously: reading a person's gesture — a construction worker waving you to stop, a cyclist signalling a turn with an arm. There are endless ways to gesture, you can't collect that many examples, and the cost of getting it wrong is a life.
- Medicine. Can: spot pneumonia on an X-ray when there are thousands of images labelled "sick / healthy". Can't: learn a diagnosis from a dozen pictures in a textbook, the way a medical student does. A person generalises from a few examples; AI needs them in bulk.
The underrated weakness: poor generalisation
Here's a boundary almost nobody talks about. AI is strong on data similar to what it saw in training, and fragile on anything unlike it. Take a model trained on one clinic's clean X-rays, move it to another where the images are shot at a different angle on different equipment with more noise, and the accuracy drops noticeably. A real doctor adapts to a new hospital in a day; AI can't.
Remember the principle: AI doesn't "understand" the task, it fits an answer to familiar examples — change the examples and the confidence falls apart. This follows straight on from the data lesson: what matters isn't only the quality of the data but how much it resembles where the model will actually work.
Why it looks like AI can do anything
Be careful with the news. Only AI's wins make headlines: beat the champion, passed the exam, painted a masterpiece. The failures — hundreds of quiet projects where AI never took off — barely get published. That skew is where the illusion of omnipotence comes from. A realistic picture means keeping the other half in mind too — the half nobody mentions.
Hold on a second: if AI could reliably predict tomorrow's share price, would anyone be selling you access to it for a couple of dollars a month?
Of course not — the owner would quietly get rich instead. Every "AI will predict the market" promise breaks on that simple thought. And having found no pattern, the model does what the hallucination lesson describes — hands you a confident, plausible forecast with nothing behind it.
Four questions: should you hand this to AI?
Before you give a task to the machine, run it through four questions:
- Is it a "one-second" task, and are there lots of examples? Perception and generation with plenty of data — green zone.
- Does the data resemble where the AI will actually work? On anything unlike it, AI is fragile.
- What does being wrong cost, and can I check the result? High cost with no way to check — stop.
- Does the AI have the data it needs? Anything recent, private or very niche (remember the knowledge cutoff) it may never have seen.
Green light: a one-second task with plenty of similar data, a low cost of error, a checkable result. Red: a slow, consequential decision, unfamiliar data, a high cost, and no way to verify.
Do this now
An exercise with no tools — just your head and a sheet of paper. Write down five real tasks from your week and run each through the one-second test and the four questions from this lesson. Mark which ones you'd hand to AI without worrying (recognise, summarise, sketch a draft) and which you'd keep (decisions where being wrong is costly, fresh data, accountability, unfamiliar conditions). That list is your personal "delegate / keep" map.
The mechanism and the boundaries are clear now — time to take the controls, which is what the next module is for: first we'll look at examples for your own line of work, then write your first real prompt.
Short questions on the lesson — with an explanation for every answer.