A December 2024 paper by Peter Henderson and Mark Lemley suggests that the typical contractual / license terms (example: OpenAI ToS) associated with AI model releases or services may not mean much.
The normal tools a company would use to enforce terms are (1) copyright; and (2) contract. But both may have issues.
First, copyright. The paper suggests that neither AI model weights nor AI model outputs can be copyrighted. And if there’s no copyright, there’s nothing to use as leverage for enforcement of terms of service.
The paper is on solid ground with respect to the lack of copyrightability of AI model outputs: most academics, the Copyright Office, and courts so far all agree that model outputs are not the product of human creativity and therefore cannot be copyrighted. Thus, if model owners try to control the use of outputs by copyright (e.g., a copyright license only if you follow our terms), this is probably pointless because users don’t need a copyright license for outputs. In other words, you can’t sue users for copyright infringement if they don’t follow your terms.
But the paper also suggests that users might not need a copyright license for model weights released as open weight models. If no copyright is required for model weights, then any conditional copyright license for those weights is also pointless. But this is where the paper is on shakier ground because it fails to reckon with the significant human creativity that goes into creating model weights.
It is true that the weights themselves are largely created algorithmically and driven by external training data. But they are not randomly generated artifacts: they are organized (one might even say designed) into meaningful layers and structures within the model’s architecture. Like algorithmically generated pixel values in a 3D-rendered scene, the weights gain expressive significance from their placement within a deliberately constructed system. The combination of architecture and weights forms the expressive work. And given that copyright law requires only a “modicum of creativity,” I think the better view is that model weights are indeed copyrightable except in extreme cases where human contribution is truly absent.
Second, contract. The paper suggests that contractual remedies will be largely ineffective for at least three reasons: (1) open weight models are often shared without requiring users to agree to anything; (2) many state law contractual causes of action such as unjust enrichment and trespass to chattel will likely be preempted by copyright in this context; and (3) many companies are simply reluctant to bring such claims because “companies like Google, OpenAI, and Anthropic are likely better off if most constraints on scraping are preempted.” I would add that for most companies, it is likely easier and more effective to impose technical restrictions on abusers (i.e. blocking IP addresses or users) than suing in court.
The paper concludes that, “AI terms of service are built on a house of sand . . .” but that this is positive because “terms of service applied to users of
public websites are problematic as a general matter.”
I don’t think model terms are as wobbly as the paper suggests given the likely copyrightability of open source model weights, but the paper does a good job showing how narrow such enforcement authorities will ultimately be.
