3 ms·

LLMs are being trained on published books a direct equivalent of records. People were able to get one to reproduce over 40% of Harry Potter and the Sorcerer’s

by Retric 1y ago

LLMs are being trained on published books a direct equivalent of records. People were able to get one to reproduce over 40% of Harry Potter and the Sorcerer’s stone word for word. https://arstechnica.com/features/2025/06/study-metas-llama-3-1-can-recall-42-percent-of-the-first-harry-potter-book/ https://arstechnica.com/features/2025/06/study-metas-llama-3... There’s zero chance that happened without the book being in their training corpus. Worse, there’s significant effort put into obscuring this.

otterley 1y ago
Yes, they are trained on books, but the courts so far are largely in agreement that AI models are neither copies nor derivative works of the source materials. If you’re searching for legal protections against your works being used for model training, copyright law as written today does not appear to give you cover.
- Retric 1y ago
  Stuff can take a long time to wind their way through the court system. The worst cases already failed but many are going strong, here’s a 1.5 billion dollar win for authors. https://www.kron4.com/news/technology-ai/anthropic-copyright-lawsuit-authors/ https://www.kron4.com/news/technology-ai/anthropic-copyright...
  - otterley 1y ago
    Anthropic and the authors settled over a portion of the case involving the unauthorized copying of works that were used to train the model. Obtaining the works is a step that happens before the training has begun. “the authors alleged nearly half a million books had been illegally pirated to train AI chatbots...” Finally, a settlement isn’t a “win” from a legal perspective. It’s money exchanged for dropping the case. In almost every settlement, there’s no admission of guilt or liability.
    - Retric 1y ago
      The entire case settled, the authors aren’t going to appeal when the company can’t hand out much more than the 1.5 billion in question and the company isn’t allowed to use the works in question going forward.
      - otterley 1y ago
        Before the settlement was made, Judge Alsup found as a matter of law that the training stage constituted fair use. https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_4.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.43...
        Retric 1y ago
        A judge yes, but that’s subject to appeal. The point is it never reached that stage and never will.
        otterley 1y ago
        As an attorney, I'm trying to understand what you're getting at. His opinion, while interlocutory and not binding precedent, will be cited in future cases. And his wasn't the only one. In Kadrey v. Meta Platforms, Inc., No. 23-cv-03417 (N.D. Cal. June 25, 2025) Judge Chhabria reached the same conclusion. https://storage.courtlistener.com/recap/gov.uscourts.cand.415175/gov.uscourts.cand.415175.598.0_1.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.41... In neither case has an appeal been sought.
        Retric 1y ago
        I’m saying legally things aren’t clear cut at this point. If you’ve read Kadrey the judge says harm from competing with the output of authors would be problematic. Quite relevant for software developers suing about code generation but much harder for novelists to prove. However, the judge came to the opposite conclusion about using pirate websites to download the books in question. A new AI company that is expecting to face a large number of lawsuits and win some while losing others isn’t in a great position.

by Retric 1y ago

otterley 1y ago

Yes, they are trained on books, but the courts so far are largely in agreement that AI models are neither copies nor derivative works of the source materials. If you’re searching for legal protections against your works being used for model training, copyright law as written today does not appear to give you cover.

Retric 1y ago
Stuff can take a long time to wind their way through the court system. The worst cases already failed but many are going strong, here’s a 1.5 billion dollar win for authors. https://www.kron4.com/news/technology-ai/anthropic-copyright-lawsuit-authors/ https://www.kron4.com/news/technology-ai/anthropic-copyright...
- otterley 1y ago
  Anthropic and the authors settled over a portion of the case involving the unauthorized copying of works that were used to train the model. Obtaining the works is a step that happens before the training has begun. “the authors alleged nearly half a million books had been illegally pirated to train AI chatbots...” Finally, a settlement isn’t a “win” from a legal perspective. It’s money exchanged for dropping the case. In almost every settlement, there’s no admission of guilt or liability.
  - Retric 1y ago
    The entire case settled, the authors aren’t going to appeal when the company can’t hand out much more than the 1.5 billion in question and the company isn’t allowed to use the works in question going forward.
    - otterley 1y ago
      Before the settlement was made, Judge Alsup found as a matter of law that the training stage constituted fair use. https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0_4.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.43...
      - Retric 1y ago
        A judge yes, but that’s subject to appeal. The point is it never reached that stage and never will.
        otterley 1y ago
        As an attorney, I'm trying to understand what you're getting at. His opinion, while interlocutory and not binding precedent, will be cited in future cases. And his wasn't the only one. In Kadrey v. Meta Platforms, Inc., No. 23-cv-03417 (N.D. Cal. June 25, 2025) Judge Chhabria reached the same conclusion. https://storage.courtlistener.com/recap/gov.uscourts.cand.415175/gov.uscourts.cand.415175.598.0_1.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.41... In neither case has an appeal been sought.
        Retric 1y ago
        I’m saying legally things aren’t clear cut at this point. If you’ve read Kadrey the judge says harm from competing with the output of authors would be problematic. Quite relevant for software developers suing about code generation but much harder for novelists to prove. However, the judge came to the opposite conclusion about using pirate websites to download the books in question. A new AI company that is expecting to face a large number of lawsuits and win some while losing others isn’t in a great position.