“By clustering piles of similar things together in a giant n-dimensional space, they pretty much automatically become synonym and pastische machines, regurgitating a lot of words with slight paraphrases while adding conceptually little, and understanding even less.”
Our reading, the sentence that was checked:
Because LLMs cluster similar things in a high-dimensional space, they become synonym and pastiche machines that regurgitate with slight paraphrase, adding little and understanding less.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Partial Regurgitation and how LLMs really work, paragraph 13.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-partial-regurgitation-and-how-llms-97b315-a2
We recorded 4 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2024-05-23. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- By clustering piles of similar things together in a giant n-dimensional space
- pretty much automatically become synonym and pastische machines
- regurgitating a lot of words with slight paraphrases
- adding conceptually little, and understanding even less
Our interpretation
- The claim concerns LLMs available when the post appeared, rather than a particular later model.
- The plain reading is that similarity-based internal representations cause substantial close paraphrasing, with limited ability to apply concepts beyond the source material.
- I interpret understanding behaviorally: applying relevant information correctly across unfamiliar contexts. This does not test subjective experience.
- Evidence that memorization sometimes occurs would establish only part of the claim. Establishing its causal explanation requires controlled interventions and tests that distinguish copying from generalization.
- No numerical meaning is supplied for 'a lot' or 'conceptually little'; I do not impose a percentage.
Counts as supported if: Controlled studies of relevant LLMs show that grouping similar material in their representations causes substantial close paraphrasing of source text, alongside limited conceptual transfer, while separating this mechanism from repetition in training data and material supplied in the prompt.
Counts as contradicted if: Controlled studies directly test the proposed mechanism and show that representation similarity does not explain the claimed paraphrasing, or demonstrate broad conceptual transfer incompatible with the claim's characterization of these models.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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
The 2022 memorization study, revised March 6, 2023, finds that verbatim recall increases with model capacity, training duplication and prompt length [e1]. The 2022 factual-editing study, revised January 13, 2023, reports generalization after edits to factual associations, but does not establish broad conceptual transfer [e2]. The January 29, 2026 study finds memorization from similar sequences and describes this memorization as predominantly syntactic rather than semantic [e3]. These findings establish memorization and some factual generalization, but do not settle whether grouping similar material in high-dimensional representations causes substantial slight paraphrasing with little conceptual contribution and limited understanding [e1, e2, e3].
Limitations
- Verbatim recall and memorization from similar sequences do not establish how often models produce close paraphrases in ordinary use [e1, e3].
- The studies do not establish or directly refute the proposed link between representation similarity, paraphrasing and limited conceptual transfer [e1, e2, e3].
- Factual-editing generalization does not measure understanding across unfamiliar contexts [e2].
- The later study does not establish the overall conceptual contribution of models available in May 2024 [e3].
Evidence (3 items)
Context only: arXiv,
LMs emit memorized training data.
Why it matters here: The study reports more memorization with greater model capacity, more duplicated examples, and longer prompting context. It supports the existence of regurgitation, but measures verbatim memorization rather than establishing that semantic clustering causes close paraphrases or limited understanding.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: NeurIPS; manuscript hosted by arXiv,
mediate factual predictions
Why it matters here: Causal interventions locate computations involved in factual predictions; the paper also reports generalization after editing factual associations. This weighs against a simple paraphrase-only account, but does not quantify conceptual understanding or directly refute Marcus's qualified claim that models add little.
- Publisher
- NeurIPS; manuscript hosted by arXiv
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Nature Communications,
assembling information from similar sequences
Why it matters here: The study supports memorization involving similar, nonidentical sequences, while describing that memorization as predominantly syntactic rather than semantic. It does not establish the claimed clustering mechanism or show that LLM behavior generally contributes little conceptually.
- Publisher
- Nature Communications
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
What would settle it
Controlled studies of models relevant to May 2024 that intervene on representations, measure exact and paraphrased source overlap, and test unfamiliar conceptual transfer while controlling for training duplication and supplied context. They would need to establish the proposed causal chain, show that representation similarity does not explain the claimed paraphrasing, or demonstrate broad conceptual transfer incompatible with the characterization.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Partial Regurgitation and how LLMs really work
By clustering piles of similar things together in a giant n-dimensional space, they pretty much automatically become synonym and pastische machines, regurgitating a lot of words with slight paraphrases while adding conceptually little, and understanding even less.
- Where
- paragraph 13
- Published
- 2024-05-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-partial-regurgitation-and-how-llms-97b315-a2: Current. Outcome: Unresolved. 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-partial-regurgitation-and-how-llms-97b315-a2.