A completely bullshit headline claims that Getty Images has sued the maker of Stable Diffusion for "stealing photos".

The text of the article reveals that that headline is total confusion. The case is not about theft at all; it is an allegation of copyright infringement. Both factually and legally, those two are totally different.

If someone had stolen photos from Getty, Getty would not have them any more.

So let's turn to the issue that this situation really concerns: does the output of a machine learning system infringe the copyright on items in the training set that contribute to that output?

There are possible cases where it clearly would infringe. If a substantial part of the output is very similar to one item in the training set, no stretch is required to conclude that it copies from that item.

However, people don't use machine learning system intending to get a part or a slightly modified version of some existing work. The aim is to mix, seamlessly, little bits of many training items. The items that play a role are more like artistic influences than like samples.

To find these to be copyright infringement would be disastrous to the creativity that copyright is nominally intended to promote.

The main purposes of copyright today is to keep some big companies rolling in dough, and any effect on artists is for politicians merely an excuse. For us, however, the question of what copyright law should say is mainly how to promote the arts without interfering with users' freedom.

I do not use Stable Diffusion, ChatGPT, or anything like them that exists now, because they don't respect the user's freedom.

It is a nonfree program that users can't even run, because users can't get the program's source code, or even its compiled executable. All you could possibly do with it is to identify yourself to the owner's server and send it some input data for your dossier. Then it sends back the output, over the net.

This is a manner of making a program available for usage that tramples users' freedom even worse than ordinary proprietary software. We call it SaaSS (Service as a Software Substitute) and I reject it, just as I reject nonfree executable software or source code under a nonfree license — for my freedom's sake.