Note–not specifically about the openness, but interesting news:
Anthropic won on a fair use argument regarding copyrighted materials in training data. The judge wrote, “Authors’ complaint is no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works.”
Then Meta had a smaller win just days later–except the judge said it wasn’t that the use was lawful, but that the plaintiffs had made the wrong argument.
They both still face the problem that the copyrighted materials in question were pirated. Anthropic’s is going to court, and Meta is going to attempt to discuss it first.
I’m struck by the fact that two lawsuits, filed almost simultaneously in the same U S. District Court for the Northern District of California, reached fundamentally different conclusions about whether AI training qualifies as fair use.
In Bartz v. Anthropic (Judge Alsup), the court declared that “AI training is exceedingly transformative,” treating the first fair-use factor (purpose and character) as decisive. By contrast, in Kadrey v. Meta (Judge Chhabria) the court placed primary weight on the fourth factor (market effect), holding that even highly transformative uses are not fair if they significantly harm the market—and, on the particular record, found for Meta only because the plaintiffs’ proof of market harm was inadequate. Judge Chhabria went so far as to criticize Bartz for “over-emphasizing transformativeness and under-emphasizing market harm,” an unusually open judicial sparring match rarely seen in Japan.
Still, Kadrey is careful to say that AI training is neither always unlawful nor always lawful. One might feel its emphasis on the fourth factor is a tad heavy-handed, yet Bartz could be faulted for leaning too heavily on the first factor. Seen together, the two opinions may be read as a rough attempt at balance.
Because AI training is intrinsically likely to be deemed transformative, the first factor generally favors AI developers. But if plaintiffs can prove that the model substitutes for their work, the fourth factor can offset that advantage. These two decisions thus illustrate the interplay between the first and fourth factors: when market impact outweighs transformativeness, fair use can fail. Litigation at that delicate tipping point is likely to proliferate in the coming years.
I think we are still in the very early days of regulation of AI’s use of content, and the existing legal toolkit is probably inadequate.
I also don’t think that copyright law will be the place that it is done in the long term: my guess is that there will have to be some specific regulation regarding training AI, or that drastic reforms of copyright law will be needed.