2025-10-09 20:46 UTC
This LLM security research from Anthropic also has important copyright implications:
https://www.anthropic.com/research/small-samples-poison
Replies (2)
-
@spoltier@qoto.org 2025-10-10 20:18
@HalvarFlake@mastodon.social I kind of see why you say that, but I'm not sure I understand ; reproducing copyrighted works is a very deterministic result, unlike this case.
-
@marver@mastodon.social 2025-11-22 18:46
@HalvarFlake@mastodon.social It's actually a pretty trivial revelation about LLMs since obviously training or fine-tuning LLMs gives you the ability to control specific behavior, even with a small set of training data (that is likely very specific in topic). The implications for copyright might be less severe than the fact that widespread reliance on LLM models that can't be properly audited for backdoors and that could be re-trained at any time might be more dangerous than most people imagine.