HN agrees with this.
HN does not agree that AI should be allowed to train on copyrighted data and says that AI companies are "stealing" other people's work.
HN does not agree that AI should be allowed to train on copyrighted data and says that AI companies are "stealing" other people's work.
Whether HN as a community has formed a consensus on this issue I don't know, but it's worth noting that different communities absolutely form opinions on issues because if you don't know that you can easily and unknowingly become susceptible to that community's bias.
This HN thread has a bit more detail, including links:
https://news.ycombinator.com/item?id=37879938
If you don't want to click through and just read the argument presented by OpenAI on this topic, you can find it here:
https://www.uspto.gov/sites/default/files/documents/OpenAI_R...
OpenAI may or may not believe in copyright, but they don't believe it applies in their case. From a different angle the US Copyright Office tends to agree, not accepting copyright registration for AI works. [2]
We may end up agreeing that AI models and their output should be open source or even public domain. However, I would not argue that on the basis of "because the makers do not believe in copyright". I think the copyright argument might be orthogonal in general.
[2] https://www.federalregister.gov/documents/2023/03/16/2023-05...
For your specific example, it depends on how inexact the "copy" actually is.
The four factors of fair use are:
1) the purpose and character of your use
2) the nature of the copyrighted work
3) the amount and substantiality of the portion taken, and
4) the effect of the use upon the potential market.
At some point your algorithm might be so lossy that the court rules that the amount and substantiality of what is taken is sufficiently low that your output may count as fair use. (eg. if the end result is something like [1] )
Say we argue for the sake of argument that -say- Stable Diffusion is some sort of lossy compression algorithm. it turns out that -on average- SD only notes down a few bits per image that it trains on. Thus it easily meets the amount and substantiality factor.
If you argue that SD is not a compression algorithm (and I really don't think it is), you may start to meet some of the other factors as well.
The analysis under EU law is a bit different, since there's no fair use per-se. But there's a separate rule for scraping which applies.
[1] https://fairuse.stanford.edu/2014/12/22/much-photo-need-alte...