This week’s musings on tech contracts…
What IP rights should customers get to outputs from generative AI? This article answers that question for both customers and gen-AI providers.
No-Contest and License
Customers often look for assignments of IP rights, but there’s really nothing to assign – at least if the provider knows what it’s doing. If the provider does assign IP, it may be giving away assets from training data or from its documentation. It needs to keep using those assets and sharing them with other customers, so giving up ownership doesn’t make sense.
Of course, if you’re the customer and you can get an IP assignment, great. But don’t despair if not. The most sensible IP terms involve a “no-contest” acknowledgement from the provider, as well as a license to provider assets in the training data. That license isn’t worth a ton, since it probably comes with a lot of caveats – again if the provider knows what it’s doing. But it does ensure that the provider has no right to block customer use of outputs.
Here’s sample language.
A. Provider does not claim ownership of intellectual property rights in any Output and does not contest any Customer claim to own such rights; in each case except to the extent that the Output includes Independent Assets. “Independent Asset” means any work of authorship in the training data that is subject to a copyright (1) owned by Provider or (2) owned by a third party and licensed to Provider with the right to sublicense to customers of the System.
B. Provider hereby grants Customer a perpetual, nonexclusive, worldwide license to reproduce, distribute, modify, publicly perform, and publicly display any Independent Asset included in an Output, with the right to sublicense each such right. The rights in the preceding sentence are granted (1) under Provider’s copyrights or, as applicable, (2) under third party copyrights, to the extent that the third party has granted Provider the necessary rights.
Nondisclosure and Disclaimers
If you’re the customer, you should also consider confidentiality terms covering outputs: typical nondisclosure language.
If you’re the provider, on the other hand, you should consider IP disclaimers – unless you have a small collection of training data and can confidently promise your customers that no one else has inconsistent IP rights. Assuming not:
Customer recognizes and agrees that any Output may include content subject to third party intellectual property rights, including rights that would block Customer’s use of such Output, and Provider offers no representation or warranty to the contrary, express or implied.
That doesn’t mean the provider can’t offer an IP indemnity. Many gen-AI providers do indemnify their customers against copyright claims … but that’s a separate topic.
Little or No Intellectual Property
Why can’t the provider assign IP? Why does the provider have so little to offer?
Gen-AI outputs have four possible components – and each comes with limited scope for grants of intellectual property.
- Assets “created” by the gen-AI (by the machine): The provider has no IP to grant. A human being can earn a patent or copyright for an asset created with “assistance” from gen-AI. But the provider isn’t contributing a human to output-generation. The human involved would be the customer employee: the person who writes the prompts and, in some cases, creates input data. The provider’s only contribution would be the AI – the non-human – which can’t earn a patent or copyright. And while trade secret or trademark rights could protect an AI output, the provider can’t give the customer those types of IP in this setting. A trade secret is economically valuable information that (a) isn’t generally known or ascertainable and (b) is subject to reasonable secrecy efforts (paraphrasing the USTA). So the best the provider can do is agree to keep outputs confidential – and agree not to claim its own trade secrets in outputs. Thus the confidentiality and disclaimer terms suggested above. Finally, trademarks comes primarily from use in commerce, which again the provider can’t give.
- Third-party-owned assets from training data: Obviously, the provider has no IP to assign. But it might have a license from the third party, with the right to grant a nonexclusive sublicense. (That’s not likely for most training data in an LLM – large language model – or other gen-AI trained on big chunks of the Internet.) Thus the license in B(2) above.
- Provider-owned content from training data (or even from customer inputs): The provider could own copyrights. But it would be foolish to assign them. A nonexclusive license makes more sense – like the one in Section B above. As for patents, the training data probably doesn’t include new inventions, but if it does, the provider certainly shouldn’t assign them. And it’s pretty risky to grant a patent license when you can’t identify the patent. That’s why the license above isn’t granted under the provider’s patents. But if you’re the customer and you can get a patent license, great. Also, if provider trade secrets or trademarks make their way into training data, something has gone wrong. Again, the provider shouldn’t license them. But again, if it does and you’re the customer, no problem.
- Customer or third party content from inputs: The provider has nothing to give.
In other words, no component of outputs includes IP the provider (a) owns and (b) can give up. But outputs could include licensable copyrights. And again, the provider can agree not to contest customer ownership of new content generated by the AI.
© 2024 by Tech Contracts Academy, LLC. All rights reserved. Thank you to Pixabay.com for great, free stock photos!