The AI industry is jubilant: In a double whammy of two decisions district judges in California ruled in favor of Meta and Anthropic on the question of AI model training and Fair Use. But the AI industry may slowly burn goodwill with the public.
Federal Judge Vince Chhabria ruled last Wednesday¹ that Meta’s decision to use the copyrighted books for generative AI training is protected under copyright law’s fair use defense. However he also cautioned that his decision might reflect the authors’ failure to litigate the case effectively more than the merits. The judge also noted that it could cut against LLM developers if outputs were proven to take over the market for original works.
But more importantly, the judge ruled that the use of artistic and literary works for LLM training was transformative because the models could be used for a purpose “distinct from creating or reading expressive works”.²
Similarly in the Anthropic case Judge Alsup ruled that LLM output could have a distinct “purpose and character” from Plaintiffs’ books and hence would fulfill this requirement of Fair Use protection.
No judgment was opinion was expressed about the legality of Meta and Anthropic using pirated collections of copyright-protected books to build a training dataset. Anthropic later changed its strategy and legally purchased the books.³
Both decisions stress the fact that model training itself is a transformative use of copyrighted material, but showed concerns that this defense might not work if the generated output is similar to the original or directly affected the creators ability to sell their works.
The question about how “transformative” the output is will become important again in the case of Disney against video generator Midjourney.⁴
Generative AI proponents argue that training a model is not different from a human reading a text and then using it to “generate” a new one. This view was partly reflected in both the defenses and the judgement. The argument, leaving the whole “AI training like humans learning” aside, seems to fall short considering the industrial scale of both the collection and processing of data as well as generation of output.
More importantly, the AI industry in question have showed nothing but disdain for the writers and researchers whose work they rely on, instead opting to treat authors simply as “data quarries” to be mined, while behind the scenes lobbying government to make copyright exceptions for LLM training and waiving transparency requirements.
This attitude is not new and reflects partly the attitude on privacy and data harvesting we have seen by social media providers in the past.
Arrogantly claiming Fair Use protection while at the same time suggesting to writers that they will become obsolete and be replaced by AI might burn through goodwill with the public where it still exists.
¹ Anthropic and Meta Decisions on Fair Use — Debevoise & Plimpton
² Meta Beats Authors’ Copyright Suit Over AI Training on Books — Bloomberg Law
³ ANDREA BARTZ, CHARLES GRAEBER, and KIRK WALLACE JOHNSON v ANTHROPIC PBC, Decision
⁴ Disney and Universal sue AI image creator Midjourney, alleging copyright infringement — TheGuardian