How does it draw something from a sentence I typed?

It starts with static and tidies up

Picture the fuzz on an old television. The tool begins with that. Then it asks itself the same question a few dozen times. What would this look like with slightly less noise in it, if it were a corner shop on a rainy Tuesday?

Each pass removes a little of the fuzz and commits to a little more of the picture. After enough passes there is no fuzz left, and what is underneath is the image. The method is called diffusion, and it is worth knowing the name because it explains the rest.

Nothing was fetched. There is no folder of corner shops that it rummaged through. It learned what corner shops tend to look like from a great many pictures, and then it made one up.

DOT / THE TRANSLATOR

I find this genuinely lovely. It is the only art form where the artist starts by making it worse.

Which is why it is strange about hands and hopeless at signs

The gist survives the process. The detail often does not. A hand is five similar things in a row, which is exactly the kind of thing that can come out as four or six without anything looking obviously wrong to the model.

Writing is worse. Letters are shapes it learned to draw, not words it learned to spell, so a shop sign comes out as something that looks like English from across the road and falls apart up close.

Both of these are getting better quickly, and that is the part worth holding on to. Every tell on this list has a shelf life. Somebody who learned the hand trick in 2023 and stopped there is now walking around with a test that passes everything.

Whose pictures did it learn from?

This is the argument, and it is not settled. These tools learned from enormous collections of images gathered from the open web, and a great many of those images were made by people who were never asked.

Some providers now train only on work they licensed, and say so. Others do not say. The UK Intellectual Property Office has been consulting on where the line sits, which tells you it is a live question rather than a closed one.

For you, practically: a picture you generate is fine for a birthday card and a risk on the side of a van. If the thing you are making earns money or carries your name, read the terms of the tool you used, and look for whether it indemnifies you.

What a picture can no longer prove

A photograph used to be weak evidence that something happened. It is now closer to no evidence at all, which is a genuine loss and worth naming rather than shrugging at.

The replacement is not a better eye. It is provenance: where a file came from, who published it, and whether anybody else saw the same thing. Content Credentials is a standard for attaching that record to a file, and where it is present it helps.

Where it is absent it means nothing, because most honest photographs carry none. It can confirm. It cannot rule out.

See it for yourself

Reading about it only gets you so far. This one is written to show you the thing described above, rather than describe it again.

If you remember nothing else

• It makes pictures up rather than looking them up, which is why the same words twice give two different results.
• It is good at the mood of a scene and bad at the detail in it, especially writing.
• Every visual tell has a shelf life. Anybody relying on one they learned three years ago is relying on nothing.
• Where a picture came from is now worth more than what it looks like.