PaperlayerView exact revision on arXiv

Source-linked AI summary

Interrogating LLM design under a fair learning doctrine

Johnny Tian-Zheng Wei, Maggie Wang, Ameya Godbole, Jonathan H. Choi, Robin Jia

arXiv:2502.16290v1cs.CYcs.CL

TL;DR

Output-focused copyright analysis of LLMs is limited because substantial similarity is difficult to define algorithmically and model outputs alone do not capture every risk. The paper examines training decisions through causal and correlational analyses of Pythia and proposes a fair learning doctrine. Its analyses find little effect from slight document upweighting, while simulated dataset exclusions predict varied loss changes; the authors discuss how courts may use such evidence without treating it as an automated legal judgment.

  • Problem

    Output similarity is difficult to define algorithmically, and output-focused analysis alone is an incomplete way to assess copyright issues in LLM training.

    Source: §1 · p. 1

  • Method

    The paper deconstructs Pythia and uses causal analysis of randomized upweighting and correlational analyses of dataset overlaps and simulated ablations to assess training decisions.

    Source: §3 · p. 5, §3.2 · p. 6, §3.4 · p. 11

  • Results

    Upweighting a document did not substantially affect its memorization in Pythia, while simulated dataset exclusions predicted little loss change for most datasets but noticeable changes for some.

    Source: §3.3 · p. 8, §3.4 · p. 11

  • Takeaways & Limitations

    The authors propose that courts use evidence about training design and memorization as part of adjudication, while refining a fair learning doctrine for clearer guidance.

    Source: §4 · p. 12, §5 · p. 14

    Memorization measurements provide a quantitative framework for analyzing training choices, but are not intended to automate the legal judgment of infringement.

    Source: §3.2 · p. 6

Abstract

from arXiv · show

The current discourse on large language models (LLMs) and copyright largely takes a "behavioral" perspective, focusing on model outputs and evaluating whether they are substantially similar to training data. However, substantial similarity is difficult to define algorithmically and a narrow focus on model outputs is insufficient to address all copyright risks. In this interdisciplinary work, we take a complementary "structural" perspective and shift our focus to how LLMs are trained. We operationalize a notion of "fair learning" by measuring whether any training decision substantially affected the model's memorization. As a case study, we deconstruct Pythia, an open-source LLM, and demonstrate the use of causal and correlational analyses to make factual determinations about Pythia's training decisions. By proposing a legal standard for fair learning and connecting memorization analyses to this standard, we identify how judges may advance the goals of copyright law through adjudication. Finally, we discuss how a fair learning standard might evolve to enhance its clarity by becoming more rule-like and incorporating external technical guidelines.

1 Introduction

Output-focused analysis is an incomplete way to address LLM copyright because substantial similarity is difficult to define algorithmically and does not capture every copyright risk. This work instead examines training decisions, operationalizes fair learning through their effects on memorization, and proposes a legal doctrine for acceptable LLM design.

1 Introduction

  • The paper shifts from output similarity to a structural perspective that examines how LLMs are trained.The authors argue that similarity is difficult to define algorithmically and is too narrow a view of copyright risks.Source: §1 · pp. 1–2
  • Fair learning is operationalized by measuring whether training decisions substantially affect a model’s memorization.The paper applies this approach to Pythia, an open-source LLM, using causal analysis of upweighting and correlational analysis of dataset curation.Source: §1 · p. 2
  • The paper proposes a fair learning doctrine to clarify acceptable LLM design and identify normative outcomes courts can pursue through adjudication.It also discusses making the doctrine more rule-like and incorporating external technical guidelines to improve clarity and predictability.Source: §1 · p. 2

2 Review of U.S. copyright law

U.S. copyright law balances incentives for creators with public access, and fair use is central to disputes over LLM training. The section contrasts output-based behavioral analysis with a structural approach that examines training design and memorization.

2 Review of U.S. copyright law

  • Copyright law grants authors exclusive rights while limiting them through fair use to preserve public access; courts must balance innovation against public interests in LLM disputes.Companies argue that training on copyrighted material is fair use, and the paper leaves that normative decision to courts.Source: §2 · p. 2

2.1 Substantial similarity and fair use

  • Substantial-similarity tests compare protected expressive elements after filtering out non-expressive material, with the boundary varying across tests and media.The paper describes the abstraction-filtration-comparison test for code as one example.Source: §2.1 · p. 3
  • Fair use turns on four factors, and cases involving caching and image thumbnails show how technology design can shape the analysis.The paper says courts may consider whether a design balances copyright holders’ rights with broader public interests in new functionality.Source: §2.1 · p. 3

2.2 Behavioral and structural perspectives

  • Behavioral analysis is limited because similarity is difficult to operationalize, and output similarity alone is neither necessary nor sufficient to establish infringement.The paper notes that infringement may involve copying during preprocessing, while adversarial prompting can elicit text without establishing infringement by itself.Source: §2.2 · p. 4
  • Structuralism examines model design choices—including dataset selection, filtering, and upweighting—as factors relevant to learning and memorization.The paper presents this training-focused approach as a basis for operationalizing fair learning and relating it to judicial adjudication.Source: §2.2 · pp. 4–5

3 Deconstructing an open-source LLM

The paper deconstructs Pythia’s training decisions to assess whether they substantially affected memorization, treating this as an operational measure of “fair learning.” It combines a randomized document-level analysis with a correlational dataset-level analysis, while distinguishing their evidentiary limits.

3 Deconstructing an open-source LLM

  • Pythia [4] provides an open-source case study of how training-data curation and upweighting choices can be examined in relation to copyright doctrine.The authors caution that the case study is intended to demonstrate statistical analysis of training decisions, not produce findings generalizable to all LLMs.Source: §3 · p. 5, §3.1 · p. 5

3.2 Operationalizing “fair learning”

  • Fair learning is operationalized as whether a training decision substantially affects memorization, assessed with complementary loss, MinK%, token-accuracy, and verbatim-memorization metrics.Loss and MinK% provide memorization signals; token accuracy measures predicted tokens, while verbatim memorization requires perfect prediction of a 10-token continuation.Source: §3.2 · pp. 5–6
  • These memorization metrics also capture non-expressive material, so they upper-bound rather than isolate memorization of copyright-relevant expressive elements.Separating generalization from memorization and expressive from non-expressive elements is described as extremely difficult.Source: §3.2 · p. 6

3.3 A causal analysis at the document level

  • Upweighting a single document had little effect on Pythia’s memorization, and small amounts of upweighting did not directly imply memorization.The train/test MinK% differences were small relative to natural variation across datasets, and the pattern held across memorization metrics.Source: Figure 1 · p. 7, §3.3 · p. 8
  • The randomized train/test split supports a causal estimate of document upweighting, although possible spillover between documents may violate SUTVA.The authors consider the approximation potentially appropriate because an individual datapoint’s effect on the model is generally small.Source: §3.3 · pp. 7–8
  • Holding out a train/test split for training decisions can enable post-hoc causal analysis, including analysis of whether Llama 3’s annealing-set inclusion affected memorization.The authors recommend that developers preserve such splits to support later reporting.Source: §3.3 · p. 8

3.4 A correlational analysis at the dataset level

  • Simulated ablations predict little loss change for most datasets, but noticeable changes for FreeLaw and PubMed Central.FreeLaw’s neighbors mainly come from FreeLaw itself, and removing them predicts a large change in its ablated loss.Source: Figure 4 · p. 10, §3.4 · p. 11
  • Dataset-level ablations are simulated by regressing snippet loss on neighborhood counts, but the association is not guaranteed to have a causal interpretation.The neighborhood analysis uses Elasticsearch and BM25; its density measure correlates well with most memorization metrics except verbatim memorization.Source: §3.4 · p. 11

4 Formalizing “fair learning”

The section defines fair learning as a causal standard for whether training substantially increases memorization of copyrighted data, then considers how courts might apply and clarify it. Its proposed doctrine aims to promote transparency and responsible innovation while balancing copyright holders’ and developers’ interests.

4.1 A fair learning doctrine

  • Fair learning is a training regime that does not substantially increase memorization of copyrighted data, framing the doctrine around a causal question.The authors connect that framing to legal doctrines’ role in developing social norms.Source: §4.1 · p. 12
  • Courts could promote transparency by shifting proof burdens or drawing adverse inferences, encouraging developers to document design choices and memorization analyses.The authors suggest developers may hold train/test splits for causal analysis and release performance and memorization information.Source: §4.1 · p. 12
  • The doctrine could encourage developers to consider legal and ethical implications and develop techniques that reduce memorization and copyright risks.The authors describe this as a way to foster responsible innovation.Source: §4.1 · p. 12
  • Courts can calibrate the substantiality threshold to balance copyright holders’ and model developers’ interests in each case.The passage proposes a context-sensitive threshold as a means of maintaining equilibrium between the parties.Source: §4.1 · p. 12

4.2 Litigating fair learning

  • After a prima facie infringement case, courts may invoke fair learning to assess training decisions, but plaintiffs face limited access to training information and technical expertise.The authors propose shifting the burden to developers or drawing adverse inferences, while leveraging defendants’ expertise.Source: §4.2 · p. 13
  • Defendants would initially show that their training decisions did not substantially contribute to memorization; plaintiffs could then identify specific concerns or analytical gaps.Courts should assess evidence quality: causal analysis may suffice, while correlational findings may call for additional experiments.Source: §4.2 · p. 13
  • Applying substantiality as a facts-and-circumstances standard allows case-specific judgments but risks inconsistency, making more specific guidelines potentially desirable.Courts would weigh copyright’s broader normative goals when setting the threshold.Source: §4.2 · p. 13

4.3 Enhancing the doctrine’s clarity

  • Making the doctrine more rule-like could clarify legal boundaries, while external technical guidelines offer expertise but raise accountability, legitimacy, and capture concerns.Rules can become over- or under-inclusive or obsolete; external guidance could temper inconsistency in substantiality decisions.Source: §4.3 · pp. 13–14

5 Conclusion

The conclusion presents structural analysis of training decisions as a way to assess memorization and inform copyright doctrine. It argues that courts can draw on empirical guidance and external expertise, while accounting for risks such as regulatory capture and weak public accountability.

5 Conclusion

  • The paper operationalizes fair learning through substantial memorization and uses statistical methods to assess how design decisions affect model memorization.Its proposed doctrine highlights information asymmetry, statistical-evidence quality, and the substantiality threshold as challenges for courts.Source: §5 · p. 14
  • The authors argue that training decisions’ impact on memorization can be measured and optimized, allowing machine-learning research to offer empirical guidance on copyright-safe design.They frame copyright as the first legal frontier for LLMs and suggest their legal frameworks may inform other areas.Source: §5 · p. 14
  • External expertise may support holistic governance, but reliance on it carries risks of regulatory capture and lack of public accountability.The conclusion presents these risks alongside the opportunity to refine legal doctrine through technological analysis.Source: §5 · p. 14

A Additional results on upweighting

The appendix reports that upweighting a single document generally has little effect on Pythia’s memorization. Figure 5 presents several metrics across training and test sets, and Table 1 describes regression results for MIN20%.

A Additional results on upweighting

  • Figure 5 groups training and test metrics by dataset component, with error bars denoting 95% confidence intervals.The train–test difference represents the causal effect of upweighting a single document.Source: Figure 5 · p. 19
  • The effects of upweighting a single document on Pythia-6.9B memorization are generally small.Figure 5 reports little variation between training and test sets across mean cross-entropy loss, token-level accuracy, and verbatim memorization.Source: Figure 5 · p. 19, Table 1 · p. 20
  • Table 1 reports MIN20% regression results for Pythia-6.9B, with β1 representing the train–test difference in memorization.The caption interprets β1 as the causal effect of upweighting a document.Source: Table 1 · p. 20, A Additional results on upweighting · p. 20

A.1 Additional results on simulated ablations

The appendix describes correlations between memorization and nearest-neighbor counts, including differences by measurement level and similarity threshold. It also outlines how those correlations support simulated dataset-ablation estimates.

A.1 Additional results on simulated ablations

  • For Elasticsearch, correlations increase as the similarity threshold rises and the returned neighborhoods become smaller.The table compares dataset-level correlations across similarity-score thresholds.Source: Table 2 · p. 21
  • Dataset-level correlations between memorization metrics and nearest-neighbor counts are stronger than document-level correlations, while verbatim memorization is the weakest metric.The table uses BM25 > 50 for the neighbor counts.Source: Table 3 · p. 21
  • The simulated ablations use counterfactual neighborhood counts from Table 3 and a regression to estimate ΔY as a shift parallel to the loss–neighbor-count regression line.The resulting simulated-ablation estimates are presented in Figure 4.Source: Figure 6 · p. 22
Loading 2502.16290v1…