Translating Copyright Law into Standards for Accountable AI Training
A proposed “fair learning doctrine” shifts the copyright debate from substantial similarity detection to model training standards.
Setting expectations for behavior—and then assessing against those expectations—is a cornerstone of accountability. A recent paper published at FAccT, Interrogating LLM Design under Copyright Law, argues that for copyright violation behaviors we might be better off focusing on the training standards of LLMs rather than their output violations. This shift in perspective—from outputs to development process—offers a path for establishing technical standards that ensure model training practices meet expectations derived from legal codes.
The underlying problem addressed by the paper is that LLMs can “memorize” content that they’ve been trained on, reproducing portions of their training data verbatim. LLM developers are currently facing dozens of legal cases alleging copyright violations. The paper argues that one of the challenges facing courts is that assessing output copyright violations hinges on showing substantial similarity between the original and output. But substantial similarity is a subjective legal concept that resists algorithmic implementation, meaning that we can’t necessarily expect LLMs to be able to reliably monitor and detect whether their output meets any kind of substantial similarity legal standard. Moreover, because users might prompt a model in adversarial ways to nudge a model towards outputting a response that is a copyright violation, this muddies the water around responsibility for the violation. How much responsibility should the user have?
This paper proposes an alternative focus: instead of debating whether an output looks “too similar,” legal forums might scrutinize whether training decisions substantially increased (or decreased) the risk of memorization. The paper refers to this as a “fair learning doctrine” and the authors argue that “By setting an appropriate standard, the doctrine can incentivize design choices that align with ethical and legal norms.” In essence, this reframing would allow developers to be held accountable if they didn’t implement the standard.
The paper works through a couple of analyses using Pythia, an open-source LLM trained on The Pile [2] to offer a proof-of-concept of such training standards. In one experiment the authors show that upweighting the number of times a document appears in a training dataset doesn’t substantially affect the memorization of that document. This analysis demonstrates a method that developers might use to analyze whether their model is sensitive to this kind of upweighting. In another analysis, the authors simulate what would happen if an entire dataset (like FreeLaw or PubMed Central) were excluded from training. Here they find that overlaps in data density can affect memorization risks—suggesting the relevance of dataset curation choices.
In general, these analyses are indicative, but there needs to be additional research to really flesh out what a development standard for minimizing memorization in LLMs should look like. After sufficient research, a technical standards body such as ISO or IEEE might then formalize it and socialize it. At that stage it could be used as a benchmark for any model developer. The main contribution of the paper is that it starts building a bridge between law and model training, suggesting legally informed development standards that might one day be operationalized and used for the purposes of accountability.
References
[1] Wei, J. T.-Z., Wang, M., Godbole, A., Choi, J. & Jia, R. Interrogating LLM design under copyright law. Proc. 2025 ACM Conf. Fairness, Accountability, Transparency. 3030–3045 (2025) doi:10.1145/3715275.3732193.
[2] Gao, L. et al. The Pile: An 800GB Dataset of Diverse Text for Language Modeling. arXiv (2020) doi:10.48550/arxiv.2101.00027.