Copyright for LLM outputs typically belongs to the prompter.
No lawsuits or legal rulings have tested whether model outputs can be treated as protected IP belonging to the frontier lab. Experts like Tuhin Chakrabarty point out the absence of precedent, leaving open whether companies will pursue contract claims or keep pressing for policy intervention instead.
Gurley argues the logical next step is enforcing existing terms through the courts rather than immediately seeking new rules in Washington. Frontier labs retain options such as API monitoring or output watermarking, but the scale of alleged distillation campaigns and any resulting economic effects remain unquantified in public sources.
Many users objected to claims questioning IP theft in AI distillation as contract violations or signs of being a CCP shill.
No Digg Deeper questions have been answered for this story yet.
@natolambert It is a bad precedent because it would mean that model makes can Clawback any discovery, invention, or derived economic value from their models. Which kind of destroys all of their present valuation.
There's no legal precedent that model outputs are IP. I'd argue it also doesn't stand up conceptually. The frontier labs are under no obligation to have their APIs be available to these customers, and can protect model generations in other ways if they want to.
We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.
@natolambert the proper legal mechanism for preventing use of model outputs for profit is a non-commercial license
What IP, precisely, was stolen? Because copyright on LLM output is commonly assigned to the prompter? Or should the model providers own model output? That'd make open models the *only* viable path to use coding agents?
We support open-source AI and the innovation it unlocks. But open source is not open season on American IP. When PRC firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table.
@natolambert I'd love for Anthropic to tell their enterprise customers that Claude outputs are Anthropic intellectual property lmao
@halvarflake It's one of those things where courts SHOULDN'T be involved. It's novel enough that Congress should pass a law saying what is or isn't allowed.
@natolambert actually, i'm pretty sure it's against TOS to train on model outputs, which means its a breach of contract, which is indeed a something that you can be sued over, except that the breaching party in this case is unreachable and shielded by a nation state.
@halvarflake Among my questions is whether a model is intellectual property, because it doesn’t contain the creators intellect.
@natolambert You can’t own the IP of model outputs…. You didn’t own the IP of the training data. No license can change that.
@HsiminR @natolambert And what happens when you break a TOS? PROTIP: You are just denied further service, that's all. There's no further liability for the "distiller".
The lack of legal precedent is a function of the newness of the tech. So turning to the US constitution: "[the United States Congress shall have power] To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries." How to interpret in this case?
@natolambert They expect us to trust them on so many levels, then they can't figure out how to stop certain customers from distillation using forms of API manipulation. Makes me feel much safer.
I think we are all forgetting how the open Internet works. This issue is not a single vector, it’s a complex web of influences and thinking anything other is a trap. Technologists built tools and ecosystems without any foresight into the implications of their own technologies. They externalized the downsides at scale at every turn. They get praised for publicly admitting they had a breach because the bare minimum is a reach. The claim ‘fair use’ holds under today’s environment is theft in plain sight.
@TomBukic @natolambert i’m currently paying them $100/mo for access to that model lol
@willccbb @natolambert I don't think training an open source model counts as a commercial activity. Does it?
@kurkosdr @natolambert wrong. break a TOS == breach of contract you're liable for damages
I would argue that models' output are IP of whoever is using them at the moment. The hardest parts of Software Engineering is understanding what should and can be built, not implementation itself. In science, the breakthroughs often come from asking the right questions. It is weird that someone can use AI in their day job, supervise agents, direct them, verify their outputs, etc. yet that is all someone else's IP?
@robertgraham My personal view is: Fair use ensures I cannot publish a book and disallow Anthropic from training on it. I don't see why Anthropic should be able to disallow me training on their output, by logic of reciprocity.