Post

1. Training AI is transformative. "Exceedingly," "spectacularly," "quintessentially," "very" and "among the most transformative many of us will see in our lifetimes." That's huge. Transformative uses win 99 times out of 100 in fair use cases. And it did here.

2. Judge Alsup assumed for purposes of this MSJ opinion that "the copies are in the computer," and he didn't care. Some critics of AI training have suggested that this would be a fatal concession, but they missed the point. What matters isn't what's in the model; it's *what gets out.*

3. There's nothing unfair about AI tools creating new works in the same genre or field or whatever as their training data. If those works are new, then fair use protects their creation. Copyright protects artists from competing with pirates, not new works.

4. If you lawfully own a copy, you can digitize it to take advantage of the benefits of digital formats, within limits - copies aren't multiplied or shared. (This is the only arguably novel element of Alsup's opinion, IMO, but I find his argument compelling, as far as it goes.)

24/6/25, 9:47 pm0 reposts0 likes

5. Pirated copies are seemingly radioactive, with some caveats—e.g., for works that aren't otherwise available.

6. AI training is not a market that copyright entitles authors to exploit. As a transformative use, it's excluded from the markets that matter in the fair use analysis. That doesn't mean AI companies can't or won't make deals for content; it just means the rights at issue aren't for training.

There's much more in there, but I think those are the big takeaways. More to come!

Matthew Sag

@matthewsag.bsky.social

That was a fun way of getting across the heart of the decision.