No, I don’t think it’s ethical. What bothers me the most is less the training itself than how Meta actually got the books. Courts may decide that training a model on text is transformative – like a student learning from library books – and judges in Kadrey v. Meta and Bartz v. Anthropic went that way in 2025. But Meta didn’t use a library. According to the internal messages Reisner describes, the company looked into licensing, decided it was too slow and expensive, and downloaded millions of books from LibGen instead. Employees clearly had doubts. One wrote that torrenting from a corporate laptop “doesn’t feel right,” and the decision apparently went all the way up to Zuckerberg. When a company weighs the legitimate option and knowingly picks the shortcut, I don’t think it can call that a gray area!!
I’m also not convinced by the cost argument. Anthropic’s 1.5B settlement works out to roughly 3k per book. That puts a real price on what Meta called “unreasonably expensive,” and it suggests the shortcut only looked cheap until the bill arrived. There’s also an irony I rarely see discussed. LibGen survives largely because students in lower-income countries can’t afford textbooks, and many people defend it as a way to widen access to knowledge. Meta used a resource built for people who can’t pay so it could avoid paying when it easily could. I think the same act can be understandable for a struggling student and indefensible for one of the wealthiest companies in the world. Who is taking matters, and so does why.
As the PM for Llama, I would be balancing several pressures at once. Legally, fair use is unsettled, and the way data is acquired can create liability on its own. Financially, I’d be weighing licensing costs now vs. potential settlements later. Competitively, I’d feel pressure to keep pace with rivals who may be cutting the same corners. I’d also need to keep the trust of authors, regulators, and the developers who build on Llama. Underneath all of this is a bigger question which is will people keep writing books if their work is quietly feeding the systems that compete with them?
In practice, I’d start with a simple check. Before we use any dataset, I’d ask whether I’d be comfortable seeing our internal discussions about it in the news because that is what happened to Meta. I’d also keep a clear record of where every dataset came from and who approved it. Once a book is built into a model, it’s really hard to remove, so the careful work has to happen up front. I’d say to leadership that licensed data is an advantage not only just a cost. Competitors also can’t replicate a clean dataset overnight and it reassures business customers who worry about legal exposure. I’d push for a collective licensing pool that authors could opt into, similar to how songwriters are paid when their music plays on the radio, so individual writers are paid and not only major publishers.
For me, the hardest tension would be speed, since doing this right could mean launching later. But because Llama is released openly, whatever we train on ends up in a model that millions of people have already downloaded. That makes mistakes difficult to undo, and I think it means we owe more care not less. My question in class was, if a court decides to rule that trainingg on pirated books is legal, does that make it ethical, or should tech companies still be held to a higher standard than the law requires?
