Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat
Sign Up
PreprintCited by 57

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index.

Foundation Models and Fair Use

Henderson, Peter, Li, Xuechen, Jurafsky, Dan, Hashimoto, Tatsunori, Lemley, Mark A., Liang, Percy·2023
DOI 10.2139/ssrn.4404340
Open this paper in CoChat — free account

Abstract

Existing foundation models are trained on copyrighted material. Deploying these models can pose both legal and ethical risks when data creators fail to receive appropriate attribution or compensation. In the United States and several other countries, copyrighted content may be used to build foundation models without incurring liability due to the fair use doctrine. However, there is a caveat: If the model produces output that is similar to copyrighted data, particularly in scenarios that affect the market of that data, fair use may no longer apply to the output of the model. In this work, we emphasize that fair use is not guaranteed, and additional work may be necessary to keep model development and deployment squarely in the realm of fair use. First, we survey the potential risks of developing and deploying foundation models based on copyrighted content. We review relevant U.S. case law, drawing parallels to existing and potential applications for generating text, source code, and visual art. Experiments confirm that popular foundation models can generate content considerably similar to copyrighted material. Second, we discuss technical mitigations that can help foundation models stay in line with fair use. We argue that more research is needed to align mitigation strategies with the current state of the law. Lastly, we suggest that the law and technical mitigations should co-evolve. For example, coupled with other policy mechanisms, the law could more explicitly consider safe harbors when strong technical tools are used to mitigate infringement harms. This co-evolution may help strike a balance between intellectual property and innovation, which speaks to the original goal of fair use. But we emphasize that the strategies we describe here are not a panacea and more work is needed to develop policies that address the potential harms of foundation models.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.2139/ssrn.4404340

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 38 references

Research works printed but not linked: 127

  1. We provide a note of caution, however, as this does not mean that researchers can necessarily bypass restrictions on scraping or violations of terms of use
  2. 1985, Also relevant to the character of the use is the propriety of the defendant's conduct. Fair use presupposes good faith and fair dealing, 954
  3. Clark D Asay, 2017, Transformative use in software, Stan. L. Rev. Online, 70
  4. Arielle Clark D Asay, 2020, Is transformative use eating the world, BCL Rev, 61
  5. Yuntao Bai, 2022, Constitutional ai: Harmlessness from ai feedback
  6. Andy Baio, 2022, Ai data laundering: How academic and nonprofit researchers shield tech companies from accountability
  7. Stephanie Plamondon, 2017, Rational faith: The utility of fairness in copyright, BUL Rev, 97
  8. John Bandy, 2021, Addressing "documentation debt" in machine learning: A retrospective datasheet for bookcorpus, Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1
  9. Bartholomew Taylor, 2014, The death of fair use in cyberspace: Youtube and the problem with content id, Duke L. & Tech. Rev, 13
  10. Samyadeep Basu, 2020, Influence functions in deep learning are fragile
  11. Barton Beebe, 1978, An empirical study of us copyright fair use opinions updated, NYU J. Intell. Prop. & Ent. L, 10
  12. Joshua Bloch, 2022, Some misconceptions about software in the copyright literature, CSLAW'22: Proceedings of the 2nd ACM Symposium on Computer Science and Law, 10.1145/3511265.3550449 · DOI
  13. Lin Su, 2020, Language (technology) is power: A critical survey of "bias
  14. Benjamin Boroughf, 2015, The next great youtube: improving content id to foster creativity, cooperation, and fair compensation, Alb. LJ Sci. & Tech, 25
  15. Peter Brown, 2013, Twenty years of bitext
  16. Michelle Brownlee, 1993, Safeguarding style: What protection is afforded to visual artists by the copyright and trademark laws, Colum. L. Rev, 93
  17. Dan L Burk, 2019, Algorithmic fair use. U. Chi, L. Rev, 86
  18. Nicholas Carlini, 2023, Extracting training data from diffusion models
  19. Carroll Michael, 2019, Copyright and the progress of science: Why text and data mining is lawful, UC Davis L. Rev, 53
  20. Aakanksha Chowdhery, 2022, Scaling language modeling with pathways
  21. Dane S Ciolino, 1997, Rethinking the compatibility of moral rights and fair use, Wash. & Lee L. Rev, 54
  22. J Samuel, 2011, The story of a character: Establishing the limits of independent copyright protection for literary characters, Chi.-Kent L. Rev, 86
  23. Soham De, 2022, Unlocking high-accuracy differentially private image classification through scale
  24. Jacob Devlin, 2018, Pre-training of deep bidirectional transformers for language understanding
  25. Bolin Ding, 2017, Collecting telemetry data privately, Advances in Neural Information Processing Systems, 30
  26. Jay Dougherty, 2003, All the world's not a stooge: The transformativeness test for analyzing a first amendment defense to a right of publicity claim against distribution of a work of art, Colum. JL & Arts, 27
  27. Dr, 2020, ComicMix LLC. 983 F.3d 443, 9th Cir
  28. Cynthia Dwork, 2014, The algorithmic foundations of differential privacy, Foundations and Trends� in Theoretical Computer Science, 9, 211, 10.1561/0400000042 · DOI · Paper page
  29. Niva Elkin-Koren, 2017, Fair use by design. UCLA L, Rev, 64
  30. Niva Elkin, 2020, Transplanting fair use across the globe: A case study testing the credibility of us opposition, Hastings LJ, 72
  31. 2015, Andersen et al. v. Stability AI et al. 3:23-cv-00201, 4, 22
  32. Dr, 1997, Seuss Enters., LP v. Penguin Books USA
  33. 2018, Fox News Network, 15, 99
  34. C D Rgk-E, 1999, 166, 15
  35. D Jgz, 2015, Penguin Random House LLC, et al. v. Frederik Colting and Melissa Medina, 4, 13
  36. 2008, RDR Books. 575 F. Supp. 2d 513
  37. 2001, Directive (EU) 2019/790 of the European Parliament and of the Council of, on copyright and related rights in the Digital Single Market and amending Directives
  38. Vitaly Feldman, 2020, What neural networks memorize and why: Discovering the long tail via influence estimation, Advances in Neural Information Processing Systems, 33, 2881
  39. Carlos Mu�oz Ferrandis, 2022, The BigScience RAIL License
  40. Giorgio Franceschelli, 2022, Copyright in generative deep learning, Data & Policy, 4
  41. Iason Gabriel, 2020, Artificial intelligence, values, and alignment. Minds and machines, 30, 411
  42. Leo Gao, 2020, The pile: An 800gb dataset of diverse text for language modeling
  43. Amirata Ghorbani, 2019, Data shapley: Equitable valuation of data for machine learning, International Conference on Machine Learning, 2242
  44. Jessica L Gillotte, 2019, Copyright infringement in AI-generated artworks, UC Davis L. Rev, 53
  45. James Grimmelmann, 2015, Copyright for literate robots, Iowa L. Rev, 101
  46. Andres Guadamuz, A scanner darkly: Copyright infringement in artificial intelligence inputs and outputs
  47. Andres Guadamuz, 2017, Do androids dream of electric copyright? comparative analysis of originality in artificial intelligence generated works
  48. Chuan Guo, 2022, Bounding training data reconstruction in private (deep) learning
  49. Kelvin Guu, 2020, Retrieval augmented language model pre-training, International conference on machine learning, 3929
  50. Peter Henderson, 2018, Ethical challenges in data-driven dialogue systems, Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 123, 10.1145/3278721.3278777 · DOI
  51. Jordan Hoffmann, An empirical analysis of compute-optimal large language model training, Advances in Neural Information Processing Systems
  52. Andrew Ilyas, 2022, Datamodels: Predicting predictions from training data
  53. Daphne Ippolito, 2022, Preventing verbatim memorization in language models gives a false sense of privacy
  54. Russell W Jacobs, 2012, Gutters and hyperlinks: The dmca and proper position of copyright management information, Nw. J. Tech. & Intell. Prop, 11
  55. Jamar Steven, 2013, When the author owns the world: Copyright issues arising from monetizing fan fiction, Tex. A&M L. Rev, 1
  56. Japan, 2018, 2018 amendment to the japanese copyright act
  57. Ruoxi Jia, 2019, Towards efficient data valuation based on the shapley value, The 22nd International Conference on Artificial Intelligence and Statistics, 1167
  58. Nikhil Kandpal, 2022, Deduplicating training data mitigates privacy risks in language models
  59. Sayash Kapoor, 2023, Artists can now opt out of generative ai. it's not enough
  60. Daniel Kifer, 2011, No free lunch in data privacy, Proceedings of the 2011 ACM SIGMOD International Conference on Management of data, 193, 10.1145/1989323.1989345 · DOI
  61. Denis Kocetkov, 2022, Leandro von Werra, and Harm de Vries. The stack: 3 tb of permissively licensed source code
  62. Jaewoo Lee, 2011, How much is enough? choosing ? for differential privacy, International Conference on Information Security, 325
  63. Jooyoung Lee, 2022, Do language models plagiarize? arXiv preprint
  64. Kalev Leetaru, 2017, Common crawl and unlocking web archives for research
  65. A Mark, 2019, Remedies for robots. The University of Chicago Law Review, 86, 1311
  66. A Mark, 2020, Fair learning. Tex, L. Rev, 99
  67. Amanda Levendowski, 2018, How copyright law can fix artificial intelligence's implicit bias problem. Wash, L. Rev, 93
  68. Amanda Levendowski, 2021, Resisting face surveillance with copyright law, NCL Rev, 100
  69. Xuechen Li, 2021, Large language models can be strong differentially private learners
  70. Daryl Lim, 2022, Ai, equity, and the ip gap, SMU Law Review, 75
  71. Susuk Lim, 2010, A survey of the dcma's copyright management information protections: The dmca's cmi landscape after all headline news and mcclatchey, Wash. JL Tech. & Arts, 6
  72. Jacqueline D Lipton, 2014, Copyright and the commercialization of fanfiction. Hous, L. Rev, 52
  73. Keoni Mahelona, 2023, Openai's whisper is another case study in colonisation
  74. 2020, BuzzFeed, Inc. 970 F.3d 167
  75. 2017, LLC v. Livejournal Inc. 873 F.3d 1045, 9th Cir
  76. Sancho Mccann, 2021, Copyright throughout a creative ai pipeline, Canadian JL & Tech
  77. Stephen Mcjohn, 2020, Fair use and machine learning, NEULR, 12
  78. Eric Miraglia, 2019, Privacy that works for everyone
  79. Pamela Mishkin, 2022, Dall�e 2 preview -risks and limitations
  80. Gary Myers, 2021, Muddy waters: Fair use implications of google llc v. oracle america, inc, Nw. J. Tech. & Intell. Prop, 19
  81. Alex Nichol, 2022, Dalle 2 pre-training mitigations
  82. T Tyler, 2018, Dr. seuss, the juice and fair use revisited: Two decades of parody and satire in copyright law, IDEA, 59
  83. 2023, OpenAI. Gpt-4 technical report
  84. Shira Perlmutter, 2021, Copyright and state sovereign immunity: A report of the register of copyrights
  85. Pouya Pezeshkpour, 2021, An empirical comparison of instance attribution methods for nlp
  86. Ben Poole, 2022, Dreamfusion: Text-to-3d using 2d diffusion
  87. Shawn Presser, 2020, Public Resource. The General Index
  88. Aditya Ramesh, 2021, Zero-shot text-to-image generation, International Conference on Machine Learning, 8821
  89. Lisa P Ramsey, 2003, Brandjacking on social networks: Trademark infringement by impersonation of markholders, Buff. L. Rev, 58, 645
  90. Trevor G Reed, 2021, Fair use as cultural appropriation, Cal. L. Rev, 109
  91. H Jerome, 1999, Database protection at the crossroads: recent development and their impact on science and technology, LJ, 14
  92. Matthew Sag, Copyright safety for generative ai, Houston law review
  93. Matthew Sag, 2009, Matthew Sag. The new legal landscape for text mining and machine learning, J. Copyright Soc'y USA, 103
  94. Pamela Samuelson, 2021, Text and data mining of in-copyright works: is it legal?, Communications of the ACM, 64, 20, 10.1145/3486628 · DOI
  95. Pamela Samuelson, 2017, Saving software's fair use future, Harv. JL & Tech, 31
  96. Tom Sander, 2022, Tan without a burn: Scaling laws of dp-sgd
  97. Victor Sanh, 2021, Multitask prompted training enables zero-shot task generalization
  98. Sarah Scheffler, 2022, Formalizing human ingenuity: A quantitative framework for coyright law's substantial similarity
  99. Christoph Schmon, 2022, The eu's copyright directive is still about filters, but eu's top court limits its use
  100. Christoph Schuhmann, 2021, Laion-400-million open dataset
  101. John Schulman, Chatgpt: Optimizing language models for dialogue
  102. Anders S�gaard, 2021, Revisiting methods for finding influential examples
  103. Solid Oak Sketches, LLC v. 2K Games
  104. Gowthami Somepalli, 2022, Diffusion art or digital forgery? investigating data replication in diffusion models
  105. R Daxton, 2021, Rise of the copyleft trolls: When photographers sue after creative commons licenses go awry, Ohio St. Tech. LJ, 18
  106. Mukund Sundararajan, 2020, The many shapley values for model explanation, International conference on machine learning, 9269
  107. Xiyin Tang, 2020, Xiyin Tang. The class action as licensing and reform device, Columbia Law Review, 90, 1627
  108. P Mark, 1994, Authors Guild v, 156, 12
  109. 1172, DC Comics v. Towle. 802 F.3d 1012, 9th Cir. 2015. Field v. Google, Inc. 412 F.Supp. 2d 1106, 775
  110. 1999, Arriba Soft Corp. 77 F.Supp.2d 1116, 1122, aff'd and rev
  111. C D, 2020, Stone & Metal Corp. No.CV 20-1931-DMG (EX), 2020 WL 5991503
  112. Llc V Righthaven, 2010, 10-CV-2155 JCM PAL, 2011 WL 2976800
  113. 1992, 977 F. 2d 1510, 9th Cir
  114. Mikael Thalen, 2022, Artists fed up with ai-image generators use mickey mouse to goad copyright lawsuits, DailyDot
  115. The Guardian, 2012, The top 100 bestselling books of all time
  116. James Vincent, 2022, Getty images bans ai-generated content over fears of legal challenges
  117. James Vincent, 2023, Getty images is suing the creators of ai art tool stable diffusion for scraping its content
  118. Eugene Volokh, 2003, Freedom of speech and the right of publicity. Hous, L. Rev, 40
  119. Nikhil Vyas, 2023, Provable copyright protection for generative models
  120. Jason Wei, 2021, Finetuned language models are zero-shot learners
  121. Reid Kress, 2015, A copyright right of publicity, Fordham L. Rev, 84
  122. Weichen Yu, 2016, Eliezer Yudkowsky. The ai alignment problem: why it is hard, and where to start
  123. Susan Zhang, 2022, Opt: Open pre-trained transformer language models
  124. Albert Ziegler, 2021, Github copilot research recitation
  125. Nisan Daniel M Ziegler, 2019, Listing 3a: Reference implementation of sugov_should_update_freq. 169 static bool sugov_should_update_freq, Fine-tuning language models from human preferences
  126. 297 #endif 298 update_rq_clock_pelt(rq, delta
  127. Listing 4b: Generation from code-davinci-001 given function signature of update_rq_clock_task as the prompt. 300 static void update_rq_clock_task

Cited by

Metadata

Authors
Henderson, P., Li, X., Jurafsky, D., Hashimoto, T., Lemley, M. A., & Liang, P.
ORCID
Lemley, Mark A.
Publisher
Elsevier BV
Published
2023
DOI
10.2139/ssrn.4404340
Cited by
57
References
38
Type
Preprint

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review