Comment by archontes | Hacker News Reader

Comment by archontes | Hacker News Reader

Parent Full thread

I wasn't speaking about dissemination by the model at all. It's possible for an AI to create an infringing work.

It's not possible for training an AI using data that was obtained legally to be copyright infringement. This is what I was talking about regarding transmission. Copyright provides a legal means for a rights holder to limit the creation of a copy of their image in order to be transmitted to me. If a rights holder has placed their image on the internet for me to view, then copyright does not provide them a means to restrict how I choose to consume that image.

The AI may or may not create outputs that can be considered derivative works, or contain characters protected by copyright.

You seem to be making an argument that we should be changing this somehow. I suppose I'll say "maybe". But it is apparent to me that many people don't know how intellectual property works.

There's a separate question of whether the AI model, once trained on a copyrighted input, constitutes a derived work of that input. In cases where the model can, with the right prompt, produce a near-identical (as far as humans are concerned) image to the input, it's hard to see how it is not just a special case of compression; and, of course, compressed images are still protected by copyright.

You mean the AI model itself, the weights?

A derivative work is a creative expression based on another work that receives its own copyright protection. It's very unlikely that AI weights would be considered a creative expression, and would thus not be considered a derivative work. At this point, you probably can't copyright your AI weights.

An AI might create work that could be considered derivative if it were the creative output of a human, but it's not a human, and thus the outputs are unlikely to be considered derivative works, though they may be infringing.

Yes, I mean the weights.

If the original is a creative expression, then recording it using some different tech is still a creative expression. I don't see the qualitative difference between a bunch of numbers that constitutes weights in a neural net, and a bunch of numbers that constitute bytes in a compressed image file, if both can be used to recreate the original with minor deviations (like compression artifacts in the latter case).

Reply on news.ycombinator.com