Image generation

Image models learned from huge collections of pictures paired with descriptions. Given a new description, they build an image that fits it, often by starting from random noise and refining it step by step until the noise looks like what you asked for.

The newest ones are also good at editing. You can upload a photo and say "remove the bin", "make it evening" or "put this logo on a mug", and the rest of the picture stays largely the same. Text inside images, once a giveaway, is now often spelled correctly.

That progress cuts both ways. A convincing fake costs nothing to make, so the big apps label their images with hidden watermarks and metadata. Those labels help, but they can be stripped, so the absence of one proves nothing.