@pluralistic@jwz@eff@mozilla It's a copyrights issue though, right? As in whether the "Fair use" exception applies as much to a billion $ company training a for-profit model as it does to a human reading an article to "learn" something?
I can't tell if you read the article, but I explain it there. That's how fair use works. Virtually no one understands fair use (hint, if you've been told that whether the defendant has a billion dollars matters, you've been misinformed), and the AI debate is filled with technical misinformation about fair use.
"I don't like this" does not cash out to "this isn't fair use," even if "this" is objectively bad.
@pluralistic@jwz@eff@mozilla Your article is focusing on memorization, but the contention is about copying and use. 0.000033% figure is not really relevant because Midjourney made copies of all the pixels from images (before analyzing them).
No, I also address transient copying. This is firmly settled as fair. Otherwise we wouldn't have search engines, you couldn't do clean-room reimplementations, and every framebuffer and stick of RAM would need a copyright license for every operation.
@pluralistic@jwz@eff@mozilla See the factor #4 again. Search engines respected robots.txt (thus were explicitly permitted) and increased the traffic to the work itself. Therefore, they did not hurt the potential market for the work.
Again, if you think the four factors are the comprehensive test for fairness, you don't understand fair use. The four factors are exemplary, not exhaustive, and the most important FU cases (many of which are precedential in the AI training question) flunk all four, most notably Betamax, the single most important tech/fair use case in US history.
Also, as noted in the article (and in today's newsletter): requiring licensing for training does nothing to resolve the labor or privacy questions.
Universal Music is making voice clones with the works that it strong-armed artists into assigning to them on unfair terms and will use them to erode those workers wages: