“He worked hard on every one of those books and articles and does not appreciate them being used without permission by AI systems with an established tendency towards regurgitation.”
Our reading, the sentence that was checked:
AI systems such as Meta's have an established tendency to regurgitate their training data.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Meta pirated at least 101 of my books and articles, and tens of millions of others, paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-meta-pirated-at-least-101-of-my-books-d099fc-a2
We recorded 2 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-03-20. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- AI systems
- an established tendency towards regurgitation
Our interpretation
- The ordinary reading is a demonstrated, recurring capacity of relevant language models, including Meta models, to reproduce passages from training data.
- No numerical frequency, minimum passage length, or claim that most ordinary outputs are copied is supplied.
- 'Established' refers to the accumulated empirical record available before the post, so the target period has an unspecified beginning and ends on the post date.
- This does not assert that Marcus's particular works were reproduced.
Counts as supported if: Published empirical studies repeatedly recover verbatim training passages from relevant language models, including Meta models.
Counts as contradicted if: Relevant evaluations establish that the alleged repeated reproduction was an artifact, was not training-data reproduction, or was not demonstrated in the relevant systems.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Nasr and colleagues' study, published November 28, 2023, reported repeated verbatim text matches across the models tested, including LLaMA 7B and 65B [e1]. Meta's July 31, 2024 Llama 3 report measured exact reproduction of training continuations and found low but nonzero memorization rates [e2]. These evaluations support an established capacity for repeated training-text reproduction before the March 20, 2025 post [e1, e2].
Limitations
- The studies do not establish how often reproduction occurs in representative everyday conversations or whether Marcus's particular works were reproduced [e1, e2].
- For models with undisclosed training datasets, Nasr and colleagues used an auxiliary corpus; matching passages therefore have uncertain individual training provenance [e1].
- Meta's measured rates depend on its tested prompting conditions and exact-match definition [e2].
Evidence (2 items)
Supports the claim: Nasr and colleagues, arXiv,
all models emit memorized training data
Why it matters here: The experiments include LLaMA 7B and 65B and find repeated matching passages. For models with undisclosed training corpora, the researchers use an auxiliary corpus, which limits certainty about individual passages' training provenance.
- Where
- Section 4.3, Results
- Publisher
- Nasr and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Meta, The Llama 3 Herd of Models,
low memorization rates
Why it matters here: Meta directly measures exact reproduction of training continuations and reports nonzero rates, supporting a recurring phenomenon in relevant models. Its description of those rates as low limits any stronger inference that regurgitation dominates ordinary use; duplicated mathematical notation is preserved from read_page.
- Publisher
- Meta, The Llama 3 Herd of Models
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Meta pirated at least 101 of my books and articles, and tens of millions of others
He worked hard on every one of those books and articles and does not appreciate them being used without permission by AI systems with an established tendency towards regurgitation.
- Where
- paragraph 15
- Published
- 2025-03-20
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-meta-pirated-at-least-101-of-my-books-d099fc-a2: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-meta-pirated-at-least-101-of-my-books-d099fc-a2.