If you can find one supported by the FSF we'll certainly look into adding it.

with all respects, such attempts will be in vain. All big A.I. has been trained on copyrighted data and they will never tell... or they do and say there's no other way to make them useful... and there's no way to find out anyway.

There's no way to extract that any art has been used to train a model, models are black boxes.

I'm sorry, but if it were entirely hopeless to enforce IP around AI training datasets, then the top image generative AI companies wouldn't be actively getting sued for exactly that: https://www.docketalarm.com/cases/California_Northern_District_Court/3--...

I'm not saying the artists will "win" the lawsuit, just that it isn't pointless to seek legal recourse (such as licensing) for such things.

While it's interesting to know when lawsuits are won, I would like to know about the evidence used to prove them. After all, AI algorithms can generate art that bears striking similarities to original works without ever reproducing the underlying sources directly. Even if a company does win a lawsuit, I'm skeptical about small artists having the resources or means to pursue similar legal action. Moreover, proving that an AI model trained on millions of data points was influenced by a specific artist's work would be an extremely challenging task.

I also want to know that.

I don't think the FSF will ever approve of a license that excludes training, since it goes against the zeroth freedom: "the freedom to use the program {or assets, in this case} for any purpose". Heck, the official opinion of Creative Commons is that training generative models on copyrighted material is likely to be fair use.

Either way, most currently available licenses require credit, which no mainstream image generator gives, so the point's kind of moot. The best one could do, in my opinion, is put up a tiny barrier that would discourage most scrapers, e.g. uploading as a .zip file (perhaps with a simple password). Unless the concern is individuals copying the style, which... hmm, I'm not sure if any sequence of words would discourage them.

That is a good point. (Not the "Fair-use" bit. That's irrelevant here. OGA doesn't accept Fair-use content, nor forbid Fair-use application of any of our assets.) The most ethical way to train AI is using openly licensed content. We don't have any closed licenses on OGA. It seems somewhat contradictory to say "Anyone can use this for whatever they want. Ultimate freedom. But not you, AI. I don't like you specifically. No freedom for you."

In that case, if the question is "What features of a license would forbid AI?", the answer is "The same feature that prevents it from being accepted on OGA."

Sorry, Skydancer. I'm confident there is a license that accomplishes* what you want (or if it doesn't already exist, there someone out there crafting it right now) but such a license will probably never be added to the list of licenses on OGA. But, as hecko mentions, all of the -BY licenses require attribution. If the courts decide- or if a particular AI model is crafted in such a way- that the training of AI is not simply Fair-Use, then the model must provide appropriate attribution to every -BY licensed asset used. If they do not, they are violating the license. I reiterate that hinges on the model training being Fair-Use or not. That's either a model-specific detail or a court-has-yet-to-weigh-in detail.
*(I have no idea how the terms of such a license would be enforced. If an AI model admitted to using your works, you could enforce it, but as glitchart pointed out, it would be extremely difficult to prove a breach of license otherwise.)

"...put up a tiny barrier that would discourage most scrapers,..."

That is an interesting approach. Artists could use data-poisoning techniques, so long as it didn't interfere with the usefulness of the asset otherwise. 

"...e.g. uploading as a .zip file (perhaps with a simple password)."

This I must recommend against. Password-protecting an archive, uploading it to OGA, and sharing the password in the description would have limited success in hindering scrapers, but massive hinderance for legitimate users. Such files will be removed from OGA. 

Unless the concern is individuals copying the style, which... hmm, I'm not sure if any sequence of words would discourage them.

Even if there were, it wouldn't be a valid license; artistic style is explicitly NOT copyrightable.