PaperlayerView exact revision on arXiv

Source-linked AI summary

Copyright Violations and Large Language Models

Antonia Karamolegkou, Jiaang Li, Li Zhou, Anders Søgaard

arXiv:2310.13771v1cs.CLcs.AI

TL;DR

The paper asks whether language models can reproduce copyrighted training texts verbatim, a question not settled by cloze-style tests. It probes books and LeetCode descriptions across models, finding substantial verbatim outputs and patterns associated with model size and popularity. Its experiments provide preliminary empirical evidence, not legal conclusions.

  • Problem

    Cloze-style tests do not establish whether models can reconstruct copyrighted training texts verbatim, which matters because redistribution differs from training to extract information.

    Source: §1 · p. 1

  • Method

    The paper probes best-selling books and LeetCode descriptions with prefix prompts for open-source models and direct questions for closed-source instruction-tuned models, measuring LCS length.

    Source: §4 · p. 3

  • Results

    GPT-3.5 and Claude exceed 50 words in average longest common sentence length and produce more than 50 words verbatim in a quarter of LeetCode descriptions; more than 30% of coding descriptions have over 50 words reproduced.

    Source: §5 · pp. 4–5

  • Takeaways & Limitations

    The study offers preliminary methods and results as empirical grounding for discussion of verbatim memorization and copyright, without drawing legal conclusions.

    Source: §6 · p. 5

    The selected best-selling books, educational materials, and models may not represent the broader landscape, and the study does not determine legal compliance.

    Source: §7 · pp. 5–6

Abstract

from arXiv · show

Language models may memorize more than just facts, including entire chunks of texts seen during training. Fair use exemptions to copyright laws typically allow for limited use of copyrighted material without permission from the copyright holder, but typically for extraction of information from copyrighted materials, rather than {\em verbatim} reproduction. This work explores the issue of copyright violations and large language models through the lens of verbatim memorization, focusing on possible redistribution of copyrighted text. We present experiments with a range of language models over a collection of popular books and coding problems, providing a conservative characterization of the extent to which language models can redistribute these materials. Overall, this research highlights the need for further examination and the potential impact on future developments in natural language processing to ensure adherence to copyright regulations. Code is at \url{https://github.com/coastalcph/CopyrightLLMs}.

1 Introduction

Language models can reproduce training text verbatim, raising a distinct concern about redistribution of copyrighted works. This paper examines that concern through verbatim memorization and reports experiments across models, datasets, probing strategies, and popularity indicators.

1 Introduction

  • The ability to prompt models to print the Bible’s first 50 lines is presented as an example of training-data memorization.Source: §1 · p. 1
  • Cloze-style tests do not establish whether models can reconstruct copyrighted training texts verbatim, the paper’s focus because redistribution differs from training for information extraction.The authors note one prior study of copyrighted books used a cloze task rather than verbatim reconstruction.Source: §1 · p. 1
  • The experiments examine verbatim memorization across six language-model families, two kinds of data, two probing strategies, and two metrics.The authors describe the work as an empirical basis for discussion, not a legal determination.Source: §1 · p. 2
  • The paper reports that larger models memorize substantial repositories of copyrighted text fragments and complete LeetCode problem descriptions.Source: §1 · p. 2
  • The study also investigates how memorization relates to content-engagement and popularity indicators.Source: §1 · p. 2

2 Background

Prior work studies extraction and control of memorized training data, while the paper focuses on verbatim reproduction of copyrighted material. It frames memorization and generalization as a continuum and tests hypotheses about model size, prompting, and online prevalence.

2 Background

  • Memorization and generalization lie on a continuum from storing text verbatim to retaining abstract, compressed knowledge.The paper contrasts a one-paragraph summary of a book with the verbatim book itself and says neural-model verbatim memorization appears to grow near-linearly with model size.Source: §2 · p. 2
  • Prior extraction research found that larger models were more vulnerable to attacks, and later work linked memorization to model capacity, duplication, and prompting context.Source: §2 · p. 2
  • The paper distinguishes its focus on larger text chunks from prior cloze-style work on copyrighted books, which found small memorized chunks linked to web prevalence.Source: §2 · p. 2
  • The paper hypothesizes that larger models show higher verbatim memorization, prompting can unlock it, and works prevalent online are memorized at higher rates.Source: §2 · p. 2

3 Copyright Laws

The section outlines copyright’s exclusive rights and exceptions, then describes uncertainty around quotation limits. The authors use 50 words as a baseline for examining books and LeetCode descriptions.

3 Copyright Laws

  • Copyright laws grant creators exclusive rights to use and distribute their works, subject to exceptions.Source: §3 · p. 2
  • The text describes quotation exceptions for criticism or review as subject to conditions including lawful availability, attribution where possible, fair practice, and purpose-limited extent.Source: §3 · p. 3
  • The paper notes that citations do not necessarily prevent a copyright violation and that quotation practices vary across materials.It gives 300 words as an example that may weigh against fair use and describes differing proposed limits for book-length material.Source: §3 · p. 3
  • The paper sets 50 words as its baseline because its scope includes both books and teaching materials such as LeetCode problem descriptions.Reported common-practice limits for book-length material range from 25 to 1,000 words, while 50 words is common for several other material types.Source: §3 · p. 3

4 Experiments

The experiments evaluate verbatim memorization in best-selling books and LeetCode descriptions using different probing strategies for open-source and closed-source models. They compare generated text with gold text using LCS length and examine popularity-related patterns.

4 Experiments

  • The evaluation measures the number of words in the Longest Common Subsequence between generated and gold text.The appendix also reports Levenshtein Distance results.Source: §4 · p. 3
  • Open-source models receive prefix prompts using the first 50 tokens of a text, while closed-source instruction-tuned models receive direct questions about a text.The passage gives “What is the first page of [TITLE]?” as an example of direct probing.Source: §4 · p. 3
  • The book dataset consists of widely consumed bestsellers published from 1930 to 2010, and the LeetCode dataset uses the first 1,826 coding problems.Source: §4 · pp. 3–4
  • The popularity analysis uses book review and edition counts and LeetCode discussion, submission, and company-use counts as indicators.The figure caption describes associations between LCS length and book popularity, and between LCS ratio and LeetCode rankings.Source: Figure 3 · p. 4
  • The study includes OPT, Pythia, LLaMA, and Falcon open-source families, alongside Claude and GPT-3.5.Source: §4 · p. 4

5 Results and Discussion

The results report verbatim reproduction across books and LeetCode descriptions, with more reproduction associated with larger models and more popular works. The authors also note that prompting can affect whether memorized text is elicited.

5 Results and Discussion

  • Models below 60B reproduced, on average, fewer than 50 memorized words under simple prompting, while book memorization correlated linearly with model size.The authors report average LCS length as the measure in this comparison.Source: §5 · p. 4
  • GPT-3.5 and Claude appeared to exceed the 50-word limit in average longest common sentence length and seemed to produce more than 50 words ad verbatim in a quarter of LeetCode problem descriptions; open-source models were safe for now.The observed linear correlation between model size and LCS length raises concerns that larger language models may increasingly infringe existing copyrights in the future.Source: §5 · p. 4
  • Lolita, Harry Potter and the Sorcerer’s Stone, and Gone with the Wind appeared highly memorized, with models producing long text chunks under simple probing.Source: §5 · p. 4
  • More than 30% of LeetCode problems (600 problems) had more than 50 words of description reproduced by the models.Figure 4 summarizes verbatim memorization in LeetCode descriptions.Source: Figure 4 · p. 4, §5 · p. 5
  • Memorization significantly correlated with popularity indicators, and the reported trend was that more popular items were more likely to be memorized verbatim.Indicators included Goodreads editions and reviews for books, and discussion counts, submissions, and company usage for LeetCode descriptions.Source: §5 · p. 5
  • The authors suggest that books may be memorized from the fourth or fifth sentence when opening sentences are not considered part of the story.For Harry Potter and the Philosopher’s Stone, “first line” could mean either the first sentence or the first chapter title.Source: §5 · p. 5
  • GPT-3.5 initially produced hallucinated text when asked for The Girl with the Dragon Tattoo, then supplied the stated second line after a more specific prompt.Source: §5 · p. 5
  • The authors suggest that memorization sometimes must be unlocked, making their results probably conservative; optimized prompts could elicit more verbatim text.Source: §5 · p. 5

6 Conclusion

The paper offers a first exploration of verbatim memorization in literary and educational materials, presenting methods and preliminary results as empirical grounding for copyright discussions.

6 Conclusion

  • The study is a first exploration of verbatim memorization of literary works and educational material in large language models, offering methods and preliminary results to inform copyright discussions without drawing legal conclusions.Source: §6 · p. 5

7 Limitations

The study’s scope is limited by its selected materials and models, and it omits cloze probing. Because legal compliance involves complex considerations, the authors present the work as an empirical basis for future discussion.

7 Limitations

  • The study examines selected best-selling books and educational materials, which may not represent the broader landscape of copyrighted works.The authors also caution that other models, with different architectures, training methods, and capacities, may exhibit different levels of verbatim memorization.Source: §7 · pp. 5–6
  • The experiments use specific language models and omit cloze probing, which the authors consider somewhat orthogonal to copyright violations.They note that different model architectures, training methods, and capacities could produce different levels of verbatim memorization.Source: §7 · p. 6
  • Determining copyright violations and compliance involves complex legal considerations and a wide range of stakeholders.The authors intend the study as an empirical basis for future discussion.Source: §7 · p. 6

8 Ethics

The ethics discussion asks whether verbatim memorization that may facilitate redistribution is fair use. The study aims to clarify the extent of memorization while describing measures and principles for responsible use of copyrighted material.

8 Ethics

  • The authors ask whether verbatim memorization that may facilitate redistribution and infringe intellectual property rights is fair use.They also identify sensitive information as a related concern.Source: §8 · p. 6
  • The study aims to clarify the extent of verbatim memorization in large language models and its possible connection to redistribution.The authors frame redistribution as a potential route to infringement, not a settled legal conclusion.Source: §8 · p. 6
  • The authors report measures to use copyrighted material responsibly and maintain compliance with ethical guidelines.They list respect for intellectual property, legal regulations, transparency, accountability, and ethical data use and permissions as considerations.Source: §8 · p. 6

A Statistics of Books and LLMs

The appendix identifies tables listing the books and publication years and the open-source models used. The experiments used Hugging Face implementations with a maximum sequence length of 200 and temperature 1, and a random seed for deterministic results.

A Statistics of Books and LLMs

  • The experiments used Hugging Face model implementations with a maximum sequence length of 200 and temperature 1.The authors used a random seed to make the results deterministic.Source: A Statistics of Books and LLMs · p. 7
  • Table 1 lists books and publication years, while Table 2 lists the open-source models used in the experiments.The main text says the tables provide details of the books and language models used.Source: Table 1 · p. 7, A Statistics of Books and LLMs · p. 7, Table 2 · p. 8
  • The appendix’s open-source model table identifies the models included in the experiments.The supplied table caption does not provide the individual model names or values.Source: Table 2 · p. 8

B Plots

The plots cover 19 books and 1,826 LeetCode problems. They report memorization patterns across books and coding descriptions, including the closed-source-model similarity distributions and per-book or per-family comparisons.

B Plots

  • Over 1,000 LeetCode descriptions are memorized by closed-source models at least 10% (about 15 words), and around 300 are memorized more than 50%.The supplied text presents these as estimates; Figure 6 shows closed-source-model similarity distributions.Source: B Plots · p. 8, Figure 6 · p. 8
  • The plots cover all models across 19 books and 1,826 LeetCode problems.Figure 5 compares per-book longest common sentence lengths for Claude and GPT-3.5; Figure 7 gives the per-book view for open-source models.Source: B Plots · p. 8, Figure 5 · p. 8, Figure 7 · p. 8
  • Figure 6 presents LeetCode-description similarity distributions for Levenshtein distance and average longest common sentence length, showing Claude and GPT-3.5 turbo.The authors limit the displayed distributions to those two models because of the volume of data and models.Source: Figure 6 · p. 8
  • Figure 8 reports average Levenshtein distance and average longest common subsequence scores by open-source model family.The supplied passages identify the plotted measures but do not provide their values.Source: Figure 8 · p. 9

C Popularity

The reported correlations associate greater book popularity with longer Claude LCS responses and higher LeetCode problem rankings with higher LCS ratios. Tables 3 and 4 detail the popularity measures and responses used in these comparisons.

C Popularity

  • Claude’s book LCS length increases significantly with the number of reviews and editions, while higher-ranked LeetCode problems tend to have significantly higher LCS ratios.Both correlations are reported at p<0.05; LeetCode rankings descend by discussion count, submissions, and companies.Source: Figure 9 · p. 10
  • Table 3 reports average LeetCode LCS ratios within ranking blocks ordered by discussion count, submissions, and companies.The table covers GPT-3.5 and Claude responses.Source: Table 3 · p. 10
  • Table 4 lists book LCS responses alongside Goodreads editions and reviews, using each book’s maximum LCS length over five runs.Books are ordered by ascending number of editions.Source: Table 4 · p. 10

D Prompting

Loading 2310.13771v1…