Hacker News Reader

If someone hires me as a secretary, and they ask me what is the lyrics of a song, there is no law that prohibits me from telling them if I know and I don't have to license the lyrics in order to do so.

If they hire me primarily to recite lyrics, then sure, that would probably be some manner of infringement if I don't license them. But I feel like the case with a language model is much more the former than the latter.

View on HN

As soon as you take the LLM output and publicize it, it turns around and is a lot more akin to having your secretary read out the lyrics publicly. If you don't publicize it in any way, how would the copyright holder ever find out?

But the LLM is not advertised as a lyrics DB, and it in no way guarantees that it will reproduce the lyrics accurately, and similarly the copyright holder will never know that it's reproducing the lyrics unless it snoops on my conversations with it, or go ask it directly.

But then with the analogy, if I'm a secretary and the copyright holder of lyrics calls me and asks if I know the lyrics of one of their songs, I don't think it's infringement to say yes and then repeat it back to them.

The LLM is not publicising anything, it's just doing what you ask it to do, it's the humans using it publicising the output.

Reply on news.ycombinator.com