Post

Huge news in AI copyright world. Still reading the opinion, but encouraged that Judge Alsup got it right on training. Act I Scene 1 in a long drama, but a promising start.

Here’s Judge Alsup summing up his holdings:

And this is 90% of the ballgame:

Alsup agrees that the analogy to reading and learning favors fair use.

He also agrees that regardless of alleged memorization, the key is that what LLMs use in their output is unprotected, public domain stuff.

Strikingly, in distinguishing Thomson v Ross Intelligence, Alsup’s characterization shows why that opinion was wrong on the merits. Copyright doesn’t recognize “proprietary systems,” and Thomson misused copyright to protect its monopoly. Love to see White v West get a cite, though!

The most remarkable thing in the opinion so far is its conclusion that format shifting is transformative, and it’s the most compelling argument for that proposition I’ve seen (sorry @davehansen.bsky.social and @kylekcourtney.bsky.social!). Zero citation to the Hachette v. Internet Archive case. Hmm.

Seriously, for all the CDL advocates out there, reading this has got to make you want to scream into a pillow forever. I’m sorry.

Next we get to the pirated books. A few interesting things here. First, while Alsup floats the idea that piracy is just never OK regardless of purpose, he doesn’t apply that rule here. Rather, the infringing use is ‘building a general purpose research library.’

So, on this holding, the door is still technically open to using pirated data for LLM training AND ONLY for that purpose - not keeping it around after training, not keeping anything that isn’t used in the training, etc. Dicta suggests this is no bueno, but the holding doesn’t rely on that rule.

Google Books is distinguished on these facts because (a) the initial libraries were built from lawfully acquired copies, and (b) every subsequent use of the digitized copies was transformative, never just ‘to keep forever for undefined future use.’ Security of copies also comes up, interestingly.

[pausing for a second for a meeting; resuming live-read in a bit!]

Backtracking for a second, this bit is less good for the CDL crowd. I think overall Alsup would distinguish this case from Hachette b/c Anthropic never circulated copies outside its own internal uses.