First GenAI "Fair Use" Decision! What It All Means
It's a "Mixed Bag" Decision for Anthropic and Individual Book Authors (& Let's Not Forget That This Judge May Have Simply Got It Wrong)
Welcome to an “emergency edition” the brAIn! This one relates to the first federal judge to rule on the issue of “fair use” in the context of a generative AI case. Previously, only one federal judge — Stephano Bibas — had ruled on “fair use” in an AI infringement context. And Bibas rejected it as a defense as a matter of law. But that was a non-generative AI case (which I continue to believe is a distinction without a difference).
Now — in generative AI case Bartz v. Anthropic — federal Judge William Alsup (Northern District of California) just ruled in favor of Anthropic, as a matter of law, that its use of properly acquired (paid for) digital versions of plaintiffs’ books for AI training purposes is a defensible “fair use.” Judge Alsup, however, strongly signaled that no “fair use” would be found in relation to Anthropic’s use of pirated libraries for training and other purposes — thereby handing both the plaintiff book authors and Anthropic the quintessential “mixed bag.”
Critical “Take-Aways” from This Decision
So what can we take out of Judge Alsup’s ruling (and its potential impacts on other “fair use” cases to come)? I’ve read the entire written opinion, and this is what I see:
(1) In his view, the use of properly acquired (paid for) copyrighted creative works — individual books in this case — for AI LLM training purposes is a transformative “fair use” (in other words, defensible).
(2) But the Judge did not give Anthropic a victory on the separate issue of Anthropic’s use of pirated versions of those books. That issue goes to trial and — based on the Judge’s ruling — it is highly likely that Anthropic will lose on the use of those pirated works for AI training purposes.
(3) Nor did Judge Alsup rule on — or even face — the separate AI copyright infringement issue on the output/display side of the equation. The book authors here did not allege that Anthropic’s AI would spit out replicas of their works. That’s a fundamentally different situation faced by Midjourney, for example, in the recent mega lawsuit by Hollywood studios Disney and NBCUniversal (which I analyzed in this earlier post) who have shown that Midjourney’s AI does spit out recognizable replicas of their famous cartoon characters and the like. Disney called Midjourney “a bottomless pit of plagiarism” when it filed its litigation.
Another Reason Why This Ruling May Be A Limited One
Getting back to point (1) above, under the Copyright law’s relevant 4-part “fair use” analysis, the fourth factor focuses on the actual or potential “market substitution” of the purported infringing works on the copyrighted ones. “Market substitution” — generally considered to be the most important factor in the “fair use” analysis — was the ground on which the U.S. Supreme Court most rejected “fair use” as a defense in its most recent analysis in the now famous Andy Warhol/Prince case (here’s my earlier analysis of that pivotal case).
Here — in the context of individual book authors — Judge Alsup found no such “market substitution.” In his words, “The copies used to train specific LLMs did not and will not displace demand for copies of Authors’ works, or not in the way that counts under the Copyright Act.”
That is a narrow interpretation of “market substitution” — and for good reason. “Market substitution” of individual books by specific authors is a fundamentally different kind of animal than the stakes at issue in other major generative AI cases. For example, in The New York Times v. OpenAI — perhaps the single most important AI infringement case — The New York Times’ entire business, and potentially the entire news-gathering industry, are at risk of being substituted out by OpenAI’s scraping of its copyrighted works (done in order to create an alternative source for news). The Times, of course, has invested vast sums of money to finance its exclusive reporting and analysis.
A Blow To Rightsholders? Yes, But Other Rulings To Come (& Judge Alsup Got It Wrong)
Although a short-term blow to authors and rights-holders in general on the critical central issue of “fair use” on properly acquired books, Judge Alsup’s decision comes as no surprise. Just a few weeks back I predicted this result due the statements he made at the relevant court hearing. Yes, it certainly isn’t what copyright owners wanted. But — at the same time — it doesn’t mean that Judge Alsup got it right.
Alsup’s ruling comes at a time when another federal judge — Judge Vince Chabbria also in the Northern District of California — has signaled he will go the other way when he makes his final decision on “fair use,” which is expected to come any time soon. And it’s important to note that federal district courts frequently reach different decisions on novel questions of law. That’s what appellate courts are for. And Judge Alsup’s decision here is sure to be appealed.
And in my view, Judge Alsup’s reasoning is highly suspect. In essence, he equated AI training on copyrighted works to humans “training” on copyrighted works — i.e., humans like me being inspired by the works of others to create our own (which has happened since the beginnings of time). I may like the chords in a hit song, for example, so “borrow” them to create my own. But that’s a false equivalency to what AI systems are doing, as I’ve written previously in one of my most widely read pieces — i.e., “No AI Copying Is Not The Same as Human ‘Copying (& Substantial Similarity Isn’t Needed).”
Ultimately, for reasons I’ve laid out time and time again, I strongly believe that the Supreme Court ultimately will rule in favor of copyright owners (and against AI developers and their defense of “fair use”) if any of these cases make it that far.
Stay tuned ….