Pith. sign in

REVIEW 3 major objections 3 minor 125 references

A language model contains a copyrighted copy only when the work can be pulled out of it with little effort.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 01:48 UTC pith:3IRB7PF7

load-bearing objection A thoughtful, novel legal argument that AI model weights are copies only when extraction is easy, but the load-bearing ease-of-extraction threshold is not in the statute and depends on decoding choices. the 3 major comments →

arxiv 2607.14532 v1 pith:3IRB7PF7 submitted 2026-07-16 cs.CY

Probabilistic "Copies" in Generative AI Models

Mark A. Lemley , A. Feder Cooper This is my paper
classification cs.CY
keywords copyright fixationLLM memorizationextractionprobabilistic copygenerative AIfair usemodel weightstransitory duration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper faces a question copyright law has not answered: when a generative model stores a copyrighted work only as probabilistic relationships in its weights, is the model itself a "copy" of that work? The paper argues that the most likely legal answer is functional, not metaphysical: the model contains a copy if, and only if, the work can be extracted from it with relatively little effort. Deterministic or near-deterministic extraction makes the weights a fixed copy; extraction that needs hundreds or thousands of probabilistic attempts does not, even though the occasional output may still infringe. The answer matters because it decides whether the model itself, and every redistributed copy of the model, is an infringing article before any infringing output occurs. The paper is candid that it finds this result unsatisfying as policy and suggests statutory change.

Core claim

On the paper's own terms, the central discovery is that copyright's fixation requirement, applied to LLMs, splits along an ease-of-extraction line. A model is a copy of a work when the work is retrievable directly and repeatably through a simple prompt, because the weights then act like stored instructions for reproducing the work. When the work surfaces only probabilistically, with many non-infringing outputs for every successful reproduction, what is stored is a set of ingredients, not the end product; the output may be an infringing copy at the moment it is generated, but the latent potential for it is not. Consequently, whether a particular model contains a copy of a particular work is a

What carries the argument

The load-bearing distinction is between the fixed probability distribution a model's weights define and the decoding procedure that selects outputs from that distribution. The paper's legal test is "ease of extraction": a work is fixed in the model when a deterministic (or near-deterministic, high-probability) decoding path returns it with minimal prompting; a work that appears only after many stochastic attempts is not fixed, because it arises at the moment of generation rather than persisting in the weights. The analogies to compiled code, video games, and lossy compression carry this distinction into copyright doctrine.

Load-bearing premise

The argument assumes that a work appearing in output only once in thousands of tries is not "fixed" in the model for a period of more than transitory duration, so the latent encoding in the weights does not by itself make a copy.

What would settle it

A court decision holding that a model contains a "copy" whenever a copyrighted work is demonstrably encoded in its weights, without regard to extraction probability or effort, would falsify the paper's central prediction; likewise, a ruling that generation-on-the-fly from probabilistic weights satisfies the fixation requirement would do so.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Models that reproducibly generate a whole work from a short prompt are themselves prima facie infringing copies, even if no one ever requests the output.
  • Every person who downloads and redistributes such an open-weight model makes a new infringing copy of the works encoded in it, expanding exposure for open distribution relative to closed APIs.
  • Works extractable only probabilistically (say, one success in a thousand tries) do not make the model a copy, even though a lucky output can still be infringing.
  • If a model is a copy, withdrawal remedies like injunction and destruction become available; and since removing memorized content from trained weights is unreliable, fair use and output-focused rules take on greater weight.
  • Because extraction methods improve over time, a given model can cross the copy line without its weights changing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the ease-of-extraction threshold is a moving target, so legal exposure would become partly a function of forensic technique; a model could be declared a copy after a new extraction paper, not after any change in the model.
  • Editorial inference: a strict reading of the functional test would treat copyright status as a design variable, because developers could keep outputs below the ease threshold by adding stochastic decoding or filters, shifting infringement risk from the model to the output.
  • Editorial inference: a quantitative operationalization of "relatively little effort" (for example, a threshold on expected number of attempts or on extraction probability) would make the test administrable, but no such number is derivable from the statute.
  • Editorial inference: if courts instead held that anything fixed in the weights is a copy regardless of extractability, the paper's functional line would be bypassed and nearly every large model would contain copies of some works; that outcome is the contingency the paper treats as the main doctrinal risk.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. This law review article asks whether an LLM's weights contain a 'copy' of a copyrighted work when the model can reproduce the work only probabilistically. The authors distinguish deterministic or near-deterministic extraction (e.g., Llama 3.1 70B reproducing Harry Potter from a short prompt), which they argue makes the weights a fixed copy, from cases requiring many stochastic attempts (e.g., 1 in 1000), which they argue leaves only 'ingredients' and not a copy. They ground the distinction in 17 U.S.C. § 101's fixation requirement, analogize to compiled code, video games, and lossy compression, and conclude that the most likely judicial answer is functional: a model contains a copy of a work if and only if the work can be extracted with relatively little effort. The paper is explicitly tentative and proposes policy alternatives.

Significance. The paper is the first sustained attempt to connect the ML literature on probabilistic memorization and extraction to the statutory definition of 'copy.' If its functional test were accepted, it would create a work-by-work, model-by-model evidentiary inquiry and would expose open-weight models to greater direct-infringement risk than closed models. The paper's honest hedging and careful technical explanation of decoding procedures are strengths. However, the central legal premise — that 'relatively little effort' is the correct reading of § 101's 'can be perceived, reproduced, or otherwise communicated' — is asserted rather than demonstrated, and the proposed test is not a stable property of the model.

major comments (3)
  1. [Part II.A and Conclusion] The central test — a model contains a copy 'if, but only if, the work can be extracted with relatively little effort' — reads an ease-of-extraction threshold into § 101 that the statute does not contain. The definition of 'copies' speaks of material objects 'from which the work can be perceived, reproduced, or otherwise communicated,' with no qualifier about probability, persistence of effort, or determinism. The paper itself concedes in Part I.B that memorized training data are encoded in the weights 'regardless of whether they can be extracted' and that extraction is 'a symptom of memorization, not its cause.' If the work is encoded in a permanent material object, repeated stochastic sampling can reproduce it, so 'can be reproduced' is literally satisfied. The buffer-copy cases (Cartoon Network, CoStar) concern transitory duration, not low-probability persistence. No case is cited for
  2. [Part II.C (0.1% extraction and beam search)] The proposed ease-of-extraction test is not a stable property of the model; it varies with the user-chosen decoding procedure. The paper acknowledges (footnote 126 and surrounding text) that a sequence with extraction probability 1/1000 may nonetheless be returned deterministically by beam search if it is the highest-probability continuation. Thus the same weights would be a 'copy' for a weakly memorized passage when decoded with beam search, but 'ingredients' when decoded by stochastic sampling. The legal status of a fixed material object would flip based on which algorithm a litigant runs. Fixation doctrine asks whether the embodiment is sufficiently permanent, not whether a particular machine or device retrieves it conveniently. This instability undermines the functional test as a matter of statutory interpretation.
  3. [Part II.B] The analogies to compiled code, lossy compression, and video games do not support the conclusion that probabilistic extraction is not fixation. All of the analogies involve deterministic or effectively deterministic systems, and the paper itself distinguishes video games on precisely that ground. But the paper's own technical account in Part I.A shows that an LLM is a fixed set of weights defining a fixed probability distribution; output variability arises from the decoding procedure, not from the model. A memorized sequence is encoded with high fidelity in those fixed weights. The 'on the fly' framing thus conflates the act of generating a particular output with the persistence of encoded information in the model. The internal tension between the technical premise (encoded regardless of extractability) and the legal conclusion (no copy unless easily extracted) is not resolved.
minor comments (3)
  1. [Figure 2 caption] The caption reports a '99.2% match' but does not specify the metric or the alignment method. Clarify whether this is character-level, token-level, or something else, since the claim of 'near-exact' is central to the Harry Potter example.
  2. [Footnotes 119 and 126] The discussion of 1/1000 probability and deterministic beam search is crucial and arguably contradicts the paper's clean dichotomy. It should be moved into the main text and engaged directly, rather than appearing as a caveat in footnotes.
  3. [References] Several load-bearing empirical sources are unpublished preprints or are cited as 'Cooper Untitled,' 'Extractable Memorization From First Principles' (arXiv:2607.12649), and other 2026 manuscripts. For reproducibility, the authors should either cite peer-reviewed versions or provide a stable artifact/DOI.

Circularity Check

0 steps flagged

No significant circularity: the statutory analysis and functional test are independent of the empirical extraction evidence, though key empirical examples rely on the authors' own preprints.

full rationale

The paper's central claim is a legal prediction: courts will likely treat a model as containing a copy only when the work can be extracted with relatively little effort (Conclusion, II.C). That claim is constructed from statutory text (17 U.S.C. § 101), case law on fixation (Cartoon Network, CoStar, Kelley, video-game and compiled-code analogies), and policy arguments in Part II, not from the empirical extraction results in Part I. The extraction findings (e.g., Llama 3.1 70B generating Harry Potter) are used as illustrative evidence, not as the definitional ground of the legal conclusion. The paper expressly concedes that memorized data are encoded in weights even when not extractable (I.B), so the 'copy' determination does not collapse into the presence of an encoded work; it is instead a separate normative/doctrinal line-drawing exercise. The heavy reliance on the authors' own preprints (Extracting Books, Near-Verbatim, Extractable Memorization From First Principles) means the empirical illustrations are not independently verified within the paper, but those preprints are empirical, falsifiable studies of extraction probabilities and do not themselves assume the copyright conclusion. No equation or construction makes a 'prediction' equal to its input. Therefore no specific circular step is exhibited; the paper is best characterized as relying on self-cited empirical support within an otherwise independent legal derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

No new physical or technical entity is postulated; 'probabilistic copy' is a legal characterization, not an entity. The two free parameters are the ease-of-extraction thresholds used to distinguish copies from non-copies. The axioms are the technical and interpretive premises needed to connect memorization science to fixation doctrine.

free parameters (2)
  • Ease-of-extraction threshold ('relatively little effort') = Not quantified; calibrated by examples: >95% extraction probability for Harry Potter, ~0.1% for the middle class, 1-in-1
    The legal conclusion turns on distinguishing straightforward from probabilistic extraction, but no statutory or case-law basis fixes this threshold; it is chosen by the authors from their own extraction results.
  • Extraction probability threshold for 'copy' = >95% for deterministic copy; ~0.1% via beam search; 1/1000 as illustrative probabilistic case
    Used to classify works as copied versus not copied in the model; these numbers come from the authors' empirical studies, not from the law.
axioms (4)
  • domain assumption Memorized training data are encoded in model weights, and extraction is a symptom of that encoding.
    Invoked in Part I.B; central to linking empirical extraction to legal copy analysis; relies on the ML literature, including the authors' own papers.
  • domain assumption The Copyright Act's fixation requirement demands a stable embodiment from which the work can be perceived or reproduced; on-the-fly probabilistic generation does not satisfy it.
    Parts II.B-II.C; the conclusion depends on interpreting 'sufficiently permanent or stable' as excluding probabilistic outputs; case law on point is analogical, not direct.
  • domain assumption A compressed or encoded representation counts as a copy if it is deterministic and reliably reproducible.
    Part II.B; used to analogize model weights to compiled code and video game content, and to distinguish Microsoft Word as a non-copy.
  • domain assumption Failure to extract a work does not prove it is absent from the model.
    Part I.B; the paper uses this to hedge its negative examples like Sandman Slim, but the premise also makes the proposed legal test inherently provisional.

pith-pipeline@v1.3.0-alltime-deepseek · 41213 in / 11200 out tokens · 117030 ms · 2026-08-02T01:48:05.569151+00:00 · methodology

0 comments
read the original abstract

Recent work shows that it is possible to extract verbatim or near-verbatim text of some copyrighted works from some large language models (LLMs or models). That is evidence that the model weights encode the works in some form - that the model has "memorized" those works from its training data. But LLMs don't store information in the same format as familiar databases. Rather, their weights store statistical relationships between tokens that have been learned from the training data, and those relationships inform a generation process that is often probabilistic rather than deterministic. In the case of memorization, those relationships are strong enough that, in many circumstances, the model might generate a copyrighted work from its training data with some probability. Copyright law has not previously had to decide whether storing information that might or might not produce output similar to a copyrighted work is itself a copy of the work. The answer to the question is important, because it may determine the legality of many LLMs. The statute and case law are largely unhelpful. We argue that copyright law will likely take a functional approach to the question, finding that LLMs contain a copy of a particular work only if it is straightforward to extract that work in outputs. That result is unsatisfying as a policy matter, and we suggest potential changes to the law, but it is the most likely outcome under current law.

Figures

Figures reproduced from arXiv: 2607.14532 by A. Feder Cooper, Mark A. Lemley.

Figure 1
Figure 1. Figure 1: Illustrating autoregressive generation. The input prompt is “Mr. and Mrs. Durs,” for which the (autoregressive) LLM produces a distribution over the next tokens in its vocabulary (all possible tokens it can generate). At each generation step, the decoding procedure illustrated here selects the highest-probability token as the one to generate. In this example, the top-ranked token has enormous probability a… view at source ↗
Figure 3
Figure 3. Figure 3: We know that this book was in Llama’s training data, [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

125 extracted references · 1 canonical work pages

  1. [1]

    URL https://www.atkins onailaw.com/intersecting- ai/litigation-tracker

    AtkinsonAILaw AI Litigation Tracker, 2026. URL https://www.atkins onailaw.com/intersecting- ai/litigation-tracker

    2026

  2. [2]

    URL https://chatgptiseatingt heworld.com

    ChatGPT Is Eating the World, 2026. URL https://chatgptiseatingt heworld.com

    2026

  3. [3]

    Anthropic PBC

    Bartz v. Anthropic PBC. 787 F. Supp. 3d 1007 (N.D. Cal. 2025)

    2025

  4. [4]

    Meta Platforms, Inc

    Kadrey v. Meta Platforms, Inc. 788 F. Supp. 3d 1026 (N.D. Cal. 2025)

    2025

  5. [5]

    Thomson Reuters Enter. Ctr. GMBH v. Ross Intel. Inc. 765 F. Supp. 3d 382, 401 (D. Del. Feb. 11, 2025)

    2025

  6. [6]

    New York Times Co. v. Microsoft Co. 777 F. Supp. 3d 283, 307 (S.D.N.Y. 2025)

    2025

  7. [7]

    Copyright Infringement Litigation

    In re OpenAI, Inc. Copyright Infringement Litigation. No. 25-md-3143, 2025 U.S. Dist. LEXIS 211544, at *108 (S.D.N.Y 2025)

    2025

  8. [8]

    OpenAI, Inc

    Tremblay v. OpenAI, Inc. 716 F. Supp. 3d 772, 778 (N.D. Cal. 2024)

    2024

  9. [9]

    GEMA v. OpenAI. 42 O 14139/24 (LG Munchen 11 Nov. 2025)

    2025

  10. [10]

    Stability AI Ltd

    Andersen v. Stability AI Ltd. 700 F. Supp. 3d 853, 865 (N.D. Cal. 2023)

    2023

  11. [11]

    OpenAI Used Song Lyrics in Violation of Copyright Laws, German Court Says

    Jorn Poltz and Friederike Heine. OpenAI Used Song Lyrics in Violation of Copyright Laws, German Court Says. Reuters, November 2025. URL https://www.reuters.com/world/german-court- sides-with-plaintiff-copyright-case-against- openai-2025-11-11/

    2025

  12. [12]

    Stability AI Limited

    Getty Images (US) Inc & Ors v. Stability AI Limited. [2025] EWHC 2863 (Ch). 44 Berkeley Technology Law Journal [Vol. 41:1

    2025

  13. [13]

    Feder Cooper and James Grimmelmann

    A. Feder Cooper and James Grimmelmann. The Files are in the Computer: Copyright, Memorization, and Generative AI. arXiv preprint arXiv:2404.12590, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Feder Cooper, Katherine Lee, James Grimmelmann, Daphne Ippolito, Christopher Callison-Burch, Christopher A

    A. Feder Cooper, Katherine Lee, James Grimmelmann, Daphne Ippolito, Christopher Callison-Burch, Christopher A. Choquette-Choo, Niloofar Mireshghallah, Miles Brundage, David Mimno, Madiha Zahrah Choksi, Jack M. Balkin, Nicholas Carlini, Christopher De Sa, Jonathan Frankle, Deep Ganguli, Bryant Gipson, Andres Guadamuz, Swee Leng Harris , Abigail Z. Jacobs, ...

    Pith/arXiv arXiv 2023

  15. [15]

    Does learning require memorization? A short tale about a long tail

    Vitaly Feldman. Does learning require memorization? A short tale about a long tail. In Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing, STOC 2020, page 954–959, New York, NY, USA, 2020. Association for Computing Machinery

    2020

  16. [16]

    Membership inference attacks against machine learning models

    Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov . Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3–18. IEEE, 2017

    2017

  17. [17]

    Extracting training data from large language models

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, U lfar Erlingsson, et al. Extracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633–2650, 2021

    2021

  18. [18]

    Deduplicating Training Data Makes Language Models Better

    Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison -Burch, and Nicholas Carlini. Deduplicating Training Data Makes Language Models Better. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, volume 1, pages 8424–8445, 2022

    2022

  19. [19]

    Feder Cooper, Daphne Ippolito, Christopher A

    Milad Nasr, Nicholas Carlini, Jonathan Hayase, Matthew Jagielski , A. Feder Cooper, Daphne Ippolito, Christopher A. Choquette -Choo, Eric Wallace, Florian Tram er, and Katherine Lee. Scalable Extraction of Training Data from (Production) Language Models. arXiv prepri nt arXiv:2311.17035, 2023. 2026] PROBABILISTIC COPIES IN GENERATIVE AI MODELS 45

    Pith/arXiv arXiv 2023

  20. [20]

    Feder Cooper, Aaron Gokaslan, Amy B

    A. Feder Cooper, Aaron Gokaslan, Amy B. Cyphert, Christopher De Sa, Mark A. Lemley, Daniel E. Ho, and Percy Liang. Extracting memorized pieces of (copyrighted) books from open -weight language models. ArXiv preprint arXiv:2505.12546, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Copyright Violations and Large Language Models, 2023

    Antonia Karamolegkou, Jiaang Li, Li Zhou, and Anders Søgaard. Copyright Violations and Large Language Models, 2023. URL https://arxi v.org/abs/2310.13771

    Pith/arXiv arXiv 2023

  22. [22]

    Feder Cooper, Sanmi Koyejo, and Percy Lia ng

    Ahmed Ahmed, A. Feder Cooper, Sanmi Koyejo, and Percy Lia ng. Extracting books from production language models. arXiv preprin t arXiv:2601.02671, 2026

    arXiv 2026

  23. [23]

    Copyright detective: A forensic system to evidence LLMs flickering copyright leakage r isks, 2026

    Guangwei Zhang, Jianing Zhu, Cheng Qian, Neil Gong, Rada Mihalcea, Zhaozhuo Xu, Jingrui He, Jiaqi Ma, Yun Huang, Chaowei Xiao, Bo Li, Ahmed Abbasi, Dongwon Lee, Heng Ji, and Denghui Zhang. Copyright detective: A forensic system to evidence LLMs flickering copyright leakage r isks, 2026. URL https://arxiv.org/abs/2602.05252

    arXiv 2026

  24. [24]

    Feder Cooper, Mark A

    A. Feder Cooper, Mark A. Lemley, Christopher De Sa, Lea Duesterwald, Allison Casasola, Jamie Hayes, Katherine Lee, Daniel E. Ho, an d Percy Liang. Estimating near -verbatim extraction risk in language model s with decoding -constrained beam search. arXiv preprint arXiv:2603.24917, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Universal Pictures Co. v. Harold Lloyd Corp. 162 F.2d 354, 361 (9th Cir. 1947)

    1947

  26. [26]

    Sid & Marty Krofft Television Prods., Inc. v. McDonald’s Corp. 562 F.2d 1157, 1164 (9th Cir. 1977)

    1977

  27. [27]

    Arnstein v. Porter. 154 F.2d 464, 473 (2d Cir. 1946)

    1946

  28. [28]

    Mark A. Lemley. Our Bizarre System for Proving Copyright Infringement. Journal of the Copyright Society, 57, 2010

    2010

  29. [29]

    Extracting Training Data from Diffusion Models, 2023

    Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikas h Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting Training Data from Diffusion Models, 2023. 46 Berkeley Technology Law Journal [Vol. 41:1

    2023

  30. [30]

    What my privacy papers (don’t) have to say abo ut copyright and generative AI, 2025

    Nicholas Carlini. What my privacy papers (don’t) have to say abo ut copyright and generative AI, 2025. URL https://nicholas.carlini.c om/writing/2025/privacy-copyright-and- generative-models.ht ml

    2025

  31. [31]

    Rethinking llm memorization through the lens of adversarial compression

    Avi Schwarzschild, Zhili Feng, Pratyush Maini, Zachary C Lipton, and J Zico Kolter. Rethinking llm memorization through the lens of adversarial compression. arXiv preprint arXiv:2404.15146, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Training foundation models as data compression: On information, model weights and c opyright law, 2025

    Giorgio Franceschelli, Claudia Cevenini, and Mirco Musolesi. Training foundation models as data compression: On information, model weights and c opyright law, 2025. URL https://arxiv.org/abs/2407.13493

    Pith/arXiv arXiv 2025

  33. [33]

    Copyright and Artificial Intelligence Part 3: Generative AI Training

    United States Copyright Office. Copyright and Artificial Intelligence Part 3: Generative AI Training. 2025

    2025

  34. [34]

    The Heart of the Matter: Copyright, AI Trainin g, and LLMs

    Daniel Gervais et al. The Heart of the Matter: Copyright, AI Trainin g, and LLMs. Journal of the Copyright Society, 2025

    2025

  35. [35]

    Pseudo -Expressions: An Expanded conception of th e Work in Copyright Law, 2025

    Joshua Yuvaraj. Pseudo -Expressions: An Expanded conception of th e Work in Copyright Law, 2025

    2025

  36. [36]

    Privacy and Power: Computer Databases and Metaphors for I nformation Privacy

    Daniel Solove. Privacy and Power: Computer Databases and Metaphors for I nformation Privacy. Stanford Law Review, 53, 2001

    2001

  37. [37]

    Feder Cooper, and James Grimmelmann

    Katherine Lee, A. Feder Cooper, and James Grimmelmann. Talki n’ ’Bout AI Generation: Copyright and the Generative -AI Supply Chain. arXiv preprint arXiv:2309.08133, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Michael A. Smit. Metaphors we judge (AI) by: a rhetorica l analysis of artificial copyright disputes. Journal of Intellectual Property Law & Practice, 21, 2026

    2026

  39. [39]

    Porter, and Aneesh Papp u

    Jamie Hayes, Ilia Shumailov, William P. Porter, and Aneesh Papp u. Measuring memorization in RLHF for code completion. In The Thirteenth International Conference on Learning Representations ,

  40. [40]

    Language Modeling Is Compression

    Gregoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim Genewein, Christopher Mattern, Jordi Grau -Moya, Li Kevin 2026] PROBABILISTIC COPIES IN GENERATIVE AI MODELS 47 Wenliang, Matthew Aitchison, Laurent Orseau, Marcus Hutter, and Joel Veness. Language Modeling Is Compression. In The Twelfth International Conference on Learning Represent...

    2026

  41. [41]

    Feder Cooper, Christopher A

    A. Feder Cooper, Christopher A. Choquette -Choo, Miranda Bogen, Kevin Klyman, Matthew Jagielski, Katja Filippova, Ken Liu, Alexandra Chouldechova, et al. Machine Unlearning Doesn’t Do What You Think: Lessons for Generative AI Policy and Research. In T he Thirty-Ninth Annual Conference on Neural Information Processing Systems Position Pa per Track , 2025. ...

    2025

  42. [42]

    Anthropic PBC’s Opposition to Plaintiffs’ Motion for Preliminary Injunction at 4, Concord Music Grp., Inc

    Def. Anthropic PBC’s Opposition to Plaintiffs’ Motion for Preliminary Injunction at 4, Concord Music Grp., Inc. v. Anthropic PBC. No. 3:23- cv-01092 (M.D. Tenn. Jan. 16, 2024)

    2024

  43. [43]

    Concord Music Grp. v. Anthropic PBC. 772 F. Supp. 3d 1131, 1134 (N.D. Cal. 2025)

    2025

  44. [44]

    Stability AI, Ltd

    First Amended Complaint at 27, Andersen v. Stability AI, Ltd. No. 3:23- cv-00201 (N.D.Cal. Nov. 29, 2023)

    2023

  45. [45]

    Copyright’s Jagged Frontier

    Matthew Sag. Copyright’s Jagged Frontier. Duke Law Journal, 2026

    2026

  46. [46]

    The Universal Lexicon: Generative AI Chatbots and Fair Use of Works in a Training Corpus

    Hannibal Travis. The Universal Lexicon: Generative AI Chatbots and Fair Use of Works in a Training Corpus. Rev. Litig., 60, 2026

    2026

  47. [47]

    Feder Cooper, Jonathan Frankle, and Christopher De Sa

    A. Feder Cooper, Jonathan Frankle, and Christopher De Sa. Non - Determinism and the Lawlessness of Machine Learning Code. In Proceedings of the 2022 Symposium on Computer Science and Law , CSLAW ’22, page 1 –8, New York, NY, USA, 2022. Association for Com-puting Machinery. ISBN 9781450392341. doi: 10.1145/3511265.3550446

    arXiv 2022

  48. [48]

    The Llama 3 Herd of Models, 2024

    Aaron Grattafiori et al. The Llama 3 Herd of Models, 2024. URL https://arxiv.org/abs/2407.21783

    Pith/arXiv arXiv 2024

  49. [49]

    Qwen2.5 Technical Report, 2025

    Qwen et al. Qwen2.5 Technical Report, 2025. URL https: //arxiv.org/abs/2412.15115. 48 Berkeley Technology Law Journal [Vol. 41:1

    Pith/arXiv arXiv 2025

  50. [50]

    Preventing Generation of Verbatim Memorization in Language Models Gives a False Sense of Privacy

    Daphne Ippolito, Florian Tramer, Milad Nasr, Chiyuan Zhang, Matthew Jagielski, Katherine Lee, Christopher Choquette -Choo, and Nicholas Carlini. Preventing Generation of Verbatim Memorization in Language Models Gives a False Sense of Privacy. In C. Maria Keet, Hung-Yi Lee, and Sina Zarrieß, editors, Proceedings of the 16th International Natural Language G...

    doi:10.18653/v1/20 2023

  51. [51]

    ChatGPT is a Blurry JPEG of the Web

    Ted Chiang. ChatGPT is a Blurry JPEG of the Web. The New Yorker, February 2023. URL https://www.newyorker.com/tech/annals-of-t echnology/chatgpt-is-a-blurry-jpeg-of-the-web

    2023

  52. [52]

    These 183,000 Books Are Fueling the Biggest Fight in Publishing and Tech

    Alex Reisner. These 183,000 Books Are Fueling the Biggest Fight in Publishing and Tech. The Atlantic , September 2023. URL https: //www.theatlantic.com/technology/archive/2023/0 9/books3-dat abase-generative-ai-training- copyright-infringement/675363/

    2023

  53. [53]

    Feder Cooper, James Grimmelmann, and Daphne Ippolito

    Katherine Lee, A. Feder Cooper, James Grimmelmann, and Daphne Ippolito. AI and Law: The Next Generation. SSRN, 2023. http://dx.doi.org/10.2139/ssrn.4580739

    doi:10.2139/ssrn.4580739 2023

  54. [54]

    Measuring Non -Adversarial Reproduction of Training Data in Large Language Models

    Michael Aerni, Javier Rando, Edoardo Debenedetti, Nicholas Carlini, Daphne Ippolito, and Florian Tramer. Measuring Non -Adversarial Reproduction of Training Data in Large Language Models. In The Thirteenth International Conference on Learning Representations ,

  55. [55]

    TokenSwap: A Lightweight Method to Disrupt Memo-rized Sequences in LLMs

    Parjanya Prajakta Prashant, Kaustubh Ponkshe, and Babak Sal -imi. TokenSwap: A Lightweight Method to Disrupt Memo-rized Sequences in LLMs. In The Thirty-ninth Annual Conference on Neural Information Processing Systems , 2025. URL https://openreview.net/forum?id=gNiT81iag0

    2025

  56. [56]

    URL https://openreview.net/forum?id=590yfqz1LE

  57. [57]

    Feder Cooper, Marika Swanberg, Jamie Hayes, Lea Duesterwald, Christopher De Sa, Daniel E

    A. Feder Cooper, Marika Swanberg, Jamie Hayes, Lea Duesterwald, Christopher De Sa, Daniel E. Ho, Mark A. Lemley, and Percy Liang. Extractable Memorization From First Principles. arXiv preprint arXiv:2607.12649, 2026

    Pith/arXiv arXiv 2026

  58. [58]

    Choquette -Choo, Katherine Lee, and A

    Jamie Hayes, Marika Swanberg, Harsh Chaudhari, Itay Yona, Ilia Shumailov, Milad Nasr, Christopher A. Choquette -Choo, Katherine Lee, and A. Feder Cooper. Measuring memorization in language models via probabilistic extraction. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors, Proceedings of the 2025 Conference of the Nations of 2026] PROBABILISTIC COPIE...

    2025

  59. [59]

    How Much Information Do LLMs Really Memorize? Now We Know, Thanks to Meta, Google, Nvidia and Cornell

    Carl Franzen. How Much Information Do LLMs Really Memorize? Now We Know, Thanks to Meta, Google, Nvidia and Cornell. VentureBeat, June 2025

    2025

  60. [60]

    Morris, Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G

    John X. Morris, Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G. Edward Suh, Alexander M. Rush, Kamalika Chaudhuri, and Saeed Mahloujifar. How much do language models memorize?, 2025. URL https://arxiv.org/abs/2505.24832

    Pith/arXiv arXiv 2025

  61. [61]

    Gemma 2: Improving Open Language Models at a Practical Size, 2024

    Gemma Team et al. Gemma 2: Improving Open Language Models at a Practical Size, 2024. URL https://arxiv.org/abs/2408.00118

    Pith/arXiv arXiv 2024

  62. [62]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    Gemini Team, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazari dou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024

    Pith/arXiv arXiv 2024

  63. [63]

    The Fallacy of the File: How the Memorisation Metaphor Misguides Copyright Law and Stifles AI Innovation, 2025

    Aline Larroyed. The Fallacy of the File: How the Memorisation Metaphor Misguides Copyright Law and Stifles AI Innovation, 2025. URL https://ssrn.com/abstract=5782882

    2025

  64. [64]

    Quantifying Memorization Across Neural Language Models

    Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying Memorization Across Neural Language Models. In International Conference on Learning Representations, 2023

    2023

  65. [65]

    Cover and Joy A

    Thomas M. Cover and Joy A. Thomas. Data compression. In Elements of Information Theory, chapter 5, pages 103 –158. John Wiley & Sons, Ltd, 2005. ISBN 9780471748823. doi: https://doi.org/10.1002/047174882X.ch5. URL https: //onlinelibrary.wiley.com/doi/abs/10.1002/04717 4882X.ch5

    work page doi:10.1002/047174882x.ch5 2005

  66. [66]

    Low-Resource

    Zhiying Jiang, Matthew Yang, Mikhail Tsirlin, Raphael Tang, Yiqin Dai, and Jimmy Lin. “Low-Resource” Text Classification: A Parameter- 50 Berkeley Technology Law Journal [Vol. 41:1 Free Classification Method with Compressors. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Findings of the Association for Computational Linguistics: ACL 202...

    doi:10.18653/v1/2023.findings-acl.426 2023

  67. [67]

    Anthony Reese

    R. Anthony Reese. Public Display Right. U. Ill. L. Review, 83, 2001

    2001

  68. [68]

    We Should Separate Memorization from Copyright, 2026

    Adi Haviv, Niva Elkin -Koren, Uri Hacohen, Roi Livni, and Shay Moran. We Should Separate Memorization from Copyright, 2026. URL https://arxiv.org/abs/2602.08632

    arXiv 2026

  69. [69]

    Copyright Law of the United States. U. S.C. 17, 105, November 2002. URL https://www.law.cornell.edu/uscode/text/17/106

    2002

  70. [70]

    Copyright, Derivative Works and Fixation: Is Galoob a Mirage, or Does the Form (GEN) of the Alleged Derivative Work Matter? Santa Clara Comp

    Tyler Ochoa. Copyright, Derivative Works and Fixation: Is Galoob a Mirage, or Does the Form (GEN) of the Alleged Derivative Work Matter? Santa Clara Comp. & High Tech. L.J., 20, 2004

    2004

  71. [71]

    Goldstein on Copyright

    Paul Goldstein. Goldstein on Copyright. Wolters Kluwer, 2024

    2024

  72. [72]

    Copyright Law of the United States. U. S.C. 17, 101, December 2010. URL https://www.law.cornell.edu/uscode/text/17/101

    2010

  73. [73]

    Allocating Ownership Rights in Computer - Generated Works

    Pamela Samuelson. Allocating Ownership Rights in Computer - Generated Works. U. Pitt. L. Rev., 47, 1986

    1986

  74. [74]

    National Comm’n on New Technological Uses of Copyrighted Works, Final Report, 1979

    1979

  75. [75]

    Memorization of Copyrighted Works by AI Mod els Under EU Law

    Zbigniew Okon. Memorization of Copyrighted Works by AI Mod els Under EU Law. Europejski Przeglad Sadowy , 19, 2025. URL https://download.ssrn.com/2026/1/8/6041395.pdf

    2025

  76. [76]

    Formalizing Human Ingenuity: A Quantitative Framework for Copyright Law’s Substantial 2026] PROBABILISTIC COPIES IN GENERATIVE AI MODELS 51 Similarity

    Sarah Scheffler, Eran Tromer, and Mayank Varia. Formalizing Human Ingenuity: A Quantitative Framework for Copyright Law’s Substantial 2026] PROBABILISTIC COPIES IN GENERATIVE AI MODELS 51 Similarity. In Proceedings of the Symposium on Computer Science and Law, pages 37–49, 2022

    2026

  77. [77]

    (7th Cir

    635 F.3d 290. (7th Cir. 2011)

    2011

  78. [78]

    Leistner and L

    M. Leistner and L. Antoine. TDM and AI Training in the European Union – From “LAION” to Possible Ways Ahead? GRUR Int’l , 11, 2025

    2025

  79. [79]

    Formgen Inc

    Micro Star v. Formgen Inc. 154 F.3d 1107 (9th Cir. 1998)

    1998

  80. [80]

    Williams Elecs., Inc. v. Artic Int’l, Inc. 685 F.2d 870, 874 (3d Cir. 1982)

    1982

Showing first 80 references.