On DeviantArt's AI Stuff

I follow a lot of artists. Many of those artists have or have had a DeviantArt account. DeviantArt (dA) did something involving AI. Let's talk about that.

Disclaimer: I use dA as a watcher/follower, not as an artist.

What did dA do? They allegedly made an AI art model/tool and used art submitted to the website to train it. They also allegedly made this an opt-out process instead of opt-in. I say "allegedy" because that's not what the official blog post I read said. What dA did do was combine two unrelated AI things into one very badly worded Twitter post that implied the above.

You can read dA's blog post here. Hopefully that link doesn't rot. I'll cover the big points here.

• dA has created a metadata tag telling image scrapers not to use an image for AI models.
  • There's no hard-enforcement on this (similar to do-not-track headers), but that would be near impossible to implement anyway.
  • Ignoring this metadata tag is a violation of ToS
• dA explicity said it: "did not, does not, and will not use deviations submitted to DeviantArt to add to third-party AI models or training sets, or on DreamUp itself."
  • DreamUp is dA's in-house art AI that was announced
• adding the opt-out tag to uploaded art was not the default, could not be done in bulk, and was not automatically applied retroactively
  • opt-out is now the default
• dA is, as far as I know and according to the official post, the only art-hosting website to do this kind of thing
  • they want more sites to adopt the metadata tag

Some extra details from an independent post by drasayer: - the in-house AI is based on Stable Diffusion's dataset, which may have dA art but will not have any added in the future or on an ongoing basis - dA requires giving credit if a specific username was used to mimic a certain style - the opt-out flag can be added retroactively (but the display of the effect is possibly bugged(?)) - the noai meta tag now allows dA to assist in the takedown process

There are lots of people in dA's official journals getting very angry over this because of the terrible timing and wording of the in-house AI announcement. Are dA's choices perfect? I don't think so - I haven't seen any indication that not tagging AI art properly is a ToS violation (and I very much think it should be). But I also think that they're mostly going about this the right way.

DA has made it very clear that most AI art is trained using data without permission, and that they are against it.