Common legal questions around LLMs?

Learn
1

I think there are a lot of things that are unclear in this area. In light of recent policy discussions for GHC I’ve contacted the HF to ask if they could facilitate getting a legal expert to shed some light on some of these issues.

This is about legal questions

I’m creating this thread to collect common questions around the legality and legal risks around use of LLMs. With the goal of the HF hopefully being able to present a legal expert with a catalogue of questions, and sharing their answers publicly with us down the line to help Maintainers avoid pitfalls and allow them to better assess the risks of whatever policy they decide to follow when it comes to LLM use.

I’m interested in the situation in the legal frameworks of the EU and the USA in particular.

If people want to discuss the morality around the current legal framework of those tools I think that is a worthwhile discussion but I encourage people to create another thread for any discussion like that to help keep this thread on topic.

Unclear legal questions:

I’m starting off with the issues I think are most pressing for informing a future GHC policy:

Soundness of the “content right waivers” from popular LLM providers.

These attempt to assign rights for generated content to users. Legal expertise about risks associated with these waivers would be very welcome, at least for the top 3-4 providers.

In particular some of these clauses seem to imply that rights might be lost retroactively due to terms of use violations in the future. Which could be problematic if output generated through the violating user has already been included in a project. Is this a real risk?

LLM output not being copyrightable

What are the consequences and potential risks when parts of a code base are not legally protected by copyright due non-copyrightable material? I assume it would simply be public domain
which might be an issue for projects that try to prohibit certain use cases of the source through licenses.

Are there unexpected consequences like changes to (lack of) liability for contributors, maintainers or organizations owning such projects? Not just copyright liability but also liability for defects for example.

Are there noteworthy interactions one should be aware of when parts of the project are not under copyright while others are?

Copyright infringement

What is the risk of becoming liable for copyright infringement in case of accidental similarity between LLM generated output and copyrighted source code or other material? I understand there is no clear answer yet. But something more concrete than “it depends” would be quite helpful.

Other common legal pitfalls?

Are there other legal pitfalls related to LLM that a legal laymen would be likely to miss?

2

We should also consider non-accidental similarity to existing code, for example because the code was regurgitated from the training set.

3

Perhaps the better phrasing would have been non-intentional similarity from the users perspective.

4

On the subject of the violation of any third-party intellectual property right, this may be of interest, if you, as “Customer”, have a commercial relationship with Anthropic:

See Sections K.1, K.3, L.3.a and L.3.b.

5

Do other providers have similar provisions?

6

I don’t know. I’ve had an opportunity to experiment with Claude (Enterprise plan), which is why the Anthropic link was to hand.

7

I guess that’s a good legal question then: should we use only LLM services from companies with good copyright provisions and which companies does that include?

Also, what do the provisions actually cover? If the case is lost, do we still need to do a lot of work to remove the infringing code?

8

Some related links: AI - hledger