Ask HN: Who is responsible/liable for an LLM Output? | Hacker News Reader

Ask HN: Who is responsible/liable for an LLM Output?

2 pointsshaburn11 comments

In what sense? Presumably the person/party/entity that publishes the content in question. The mechanics of software are responsible for LLM output, but the logistics of spreading it is usually handled by individuals.

My primary question is from a legal standpoint, partially in response to new US regulation coming online.

This can't be answered until the regulations are finished. It's hotly debated and I'm sure everyone involved will be playing the "not me" game until the last moment.

There exists legal precidents for all other forms of software and content production. Why is this different?

Going by principles and precedents, I would say the user of the LLM will be responsible. Woke activists are trying to override this and push liability onto the developers of the models so it will depend who has better lobbyists.

I don't understand how a user will be liable when they are receiving non-deterministic outputs and have no objective way to know or control what they receive.

Ignoring the fact that most online services (and free software licenses) enforce limited liability, the actual demonstrable harm of bad LLM output stems from sharing it. If you generate a harmful, hateful, incorrect or offensive LLM generation and publish it without context, you are responsible for spreading that information. The fact that someone (or something) else wrote it doesn't appear to have any meaningful difference in-context.

I don't understand how anyone but the user could be liable for misusing an AI platform. If this upcoming legislation looks any different from Section 230, I'll be shocked.

I am not referencing republishing the output results. That is a seperate act entirely. There are countless examples, particularly in regulated usecases, where threads are audited. Thus, who is liable for the output of the LLM in those threads?

Free speech is pretty strong in the US. They can probably work around many restrictions with "this is not medical/legal/financial advice" disclaimers. It's legal for Google to give search results that are completely wrong so AI is no worse.

> I am not referencing republishing the output results.

Okay, then what exactly are the harms you're referring to? Boneheaded implementations that need a scapegoat when it breaks?

If your system isn't accounting for the regular failure of AI, you haven't built an AI-ready system.

> There are countless examples, particularly in regulated usecases, where threads are audited. Thus, who is liable for the output of the LLM in those threads?

Whoever's idea it was to give AI control over that situation is the person who assumes liability. If a government audit reveals that someone is using an LLM to respond to emails, they're not going to blame OpenAI. These situations exist because motivated individuals, knowledgeably or not, create unreliable systems that they themselves are responsible for. The only realistic and effective stance is implicating offenders on an individual basis.

This entire post feels like you're trying to direct us towards some answer, but I'm not really getting where you're going. Who do you think should be responsible?

the biggest difference is that the output is automatically generated content for human consumption* at scale.

* it remains to be seen if man or machine will be the biggest consumer of the content. It is entirely possible that the internet will become dominated by AI readers and writers instead of cats and naked activities, though these will likely become largely AI generated too, maybe not cats since they provide ample content by just being themselves

Reply on news.ycombinator.com