ceoln

There are lots of open questions about (among other things) the legal status of AI art tools (Midjourney, StableDiffusion, etc.) with respect to the works that the AIs are trained on.

What might happen in the short to medium term, with the concerns of artists and writers whose works are used to produce AI tools that other people then use to create art and writing that might (for instance) cut into the demand for the works of those artists and writers? Will the legal system find that (for instance) there are copyright issues involved at all?

I list some possibilities, in roughly descending order of how likely I think they are:

• I fear that the most likely outcome is that the artists and writers are just out of luck, because some of the AI tools companies are rich and individual artists and writers, even as a class, aren’t rich.
• Second most likely, the bigger AI-makers will give some symbolic amount of money to something that will benefit some artists and writers a little and some lawyers a lot, and there will be no precedent-setting court decision.
• Less likely, after some long wrangling process, something like the Private Copying Levy might be worked out, which is sort of like that last bullet, but more codified and involving more money, and possibly a precedent that there is a copyright violation at least potentially involved.
• Even less likely, there would be some kind of opt-out process whereby a creator could indicate they didn’t want their stuff used to train AIs, and makers of AI engines would have to like re-generate their neural nets annually without the opted-art works.
• And at the bottom, perhaps fairest in some sense but also least likely, a straightforward finding that AI Engine makers, at least ones that make money, really do need the right to copy and/or prepare derivative works of the things they train their engines on. So we’d get engines trained on just public domain works, things out of copyright, things posted under sufficiently permissive licenses, things they explicitly license, and so on. I would be fine with this, myself, but I wouldn’t bet on it happening.

Time will tell, and most likely the above is wrong in significant and interesting ways. :)

(Things might get interestingly different if big record companies eventually go up against big AI makers over AI tools for music; it's interesting in itself that this hasn't happened, and there's no Stable Diffusion for music! I wonder why.)

I wrote a version of this list on my main wordpress weblog (scroll down a bit there), in the context of a developing class action lawsuit against Microsoft and GitHub, over copyright violations in the CoPilot product (which, unlike AI tools for text and art, seems to commit verbatim copying with some frequency).

But I think the basic set of possibilities applies to generative AI tools for text and art as well, so I thought I'd post an appropriately tweaked list here as well. I guess you can't comment here :) unless I've overlooked an affordance, so feel free to comment on the original wordpress post or post at me on Mastodon or whatever!