What AI can't do
You've seen a lot of what AI does by now.
Recognising photos, guessing the next word, making pictures.
Now the other side. What can't it do?
One thing before that, though.
Three and five were the hard ones
Nothing in them was wrong. That's what made them hard.
Always calm, always tidying things up for you, and never talking about itself.
But something strange happened
Talking like a person doesn't mean thinking like one.
Think back to the 1966 counsellor in 4.4. It understood not one word. People opened up to it anyway.
Talking well and thinking were different things all along.
It's only good at one job
Count up what you've met.
| Lab | Its job |
|---|---|
| 2.2 character recognition | reading handwriting |
| 3.4 the driving range | staying on the track |
| 4.2 guessing the next word | picking the next piece |
| 5.1 ordering a picture | making a picture out of mess |
None of them can do each other's job. You can't ask a driving brain to write, or a picture machine to find a route.
Learn one thing and you'll use it elsewhere. AI mostly doesn't.
Anything needing a body is harder still
It talks beautifully and can't tie a shoelace.
Odd, isn't it. It does the hard thing and not the easy one.
Picking up an object it has never seen
Catching itself when it starts to fall
Crossing a crowded room without bumping into anyone
A five-year-old does all of that. Machines still struggle.
The easier something feels to us, the harder it tends to be for a machine. What your body learned was never written down in words.
And there's one more thing
Being responsible.
All five have one thing in common
Whoever got stuck is a person, and whoever can fix it is a person.
When the AI marker gets it wrong, you can't take it up with the AI. You can't punish it, and you can't make it apologise.
When an AI gets it wrong, the responsible party isn't the AI. It's a person.
Whoever decided to use it
Whoever built it
Whoever used its answer as it came
Why knowing the limits matters
Know only the strengths and you have no idea how far you can hand something over.
4.5 gave you one test: can I check the result myself?
Here's a second. Is there someone answerable when it goes wrong?
But the AI you meet most is a different one
Not a chatbot, not a picture maker.
That's the next chapter.
Sources for this chapter
- 3 Turing, "Computing Machinery and Intelligence", Mind 59(236) (1950)
- 30 Jones & Bergen, "Large Language Models Pass the Turing Test", arXiv:2503.23674 (2025-03-31)