Hacker News Reader

Comment by tjr | Hacker News Reader

A lot of companies avoid using GPL code at all, because they don't want to even accidentally find their own proprietary code subject to being released under the GPL.

I can imagine that some similar licensing concept could be applied to copyrightable works with respect to AI training. Use a legally restricted work for AI training, and your entire AI training set is subject to free public release?

View on HN

No such licensing concept could work. The AI companies' argument is that what they're doing counts as fair use, so they don't need to abide by any license. If that argument is true, then such a thing wouldn't do anything, and if it's false, then such a thing is unnecessary, because they're already violating even the most lax licenses (e.g., MIT) by failing to provide the required attribution.

Can't one just literally say that AI is not allowed? It's not fair use when you're going directly against the license.

> It's not fair use when you're going directly against the license.

Isn't the whole point of fair use that it's stuff you can do even if the copyright holder isn't okay with it?

https://www.copyright.gov/fair-use/

Fair use can be a bit ambiguous, but I think there are definitely grounds to claim that what the AI companies are doing is not fair use.

I imagine it will have to be tried in court for anyone to say for certain. (And I imagine whatever court ruling happens would be contested. We may actually need multiple court rulings...)

i don't think anyone training AI models cares about any of that right now

If engineers cared about ethics, we would know it.

Reply on news.ycombinator.com