“LLMs don’t store things in simple databases, but when presented with enough data, their regeneration of distributed bits has demonstrably duplicated lengthy training materials multiple times, a phenomenon that is called memorization in the field.”
Our reading, the sentence that was checked:
Given enough data, LLMs have demonstrably duplicated lengthy training materials multiple times (memorization).
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument, paragraph 51 (footnote 1).
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deconstructing-geoffrey-hintons-weakest-530eb6-a1
We recorded 10 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 2024-02-05. 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)
Conditions
- when presented with enough data (qualifier, stated by the speaker)
From the speaker
- LLMs don’t store things in simple databases
- when presented with enough data
- has demonstrably duplicated lengthy training materials multiple times
- called memorization in the field
Our interpretation
- This asserts that repeated examples of substantial training-text reproduction had been demonstrated by the post date.
- Lengthy means extended passages or substantial code, rather than common short phrases. Multiple times means more than one demonstrated reproduction, not a minimum frequency in ordinary use.
- Enough data describes the training setting. It does not specify a universal data-volume threshold or assert that every model reproduces every passage.
- The wording has demonstrably duplicated refers to accumulated past demonstrations, so the target period has no stated beginning and ends on the post date.
- The database distinction concerns text generated by learned model parameters, rather than text fetched through an external retrieval system.
Counts as supported if: Primary research documents multiple extended outputs matching training material from neural language models, with methods that distinguish memorization from retrieval or accidental phrase overlap.
Counts as contradicted if: The relevant demonstrations are invalidated, or their apparent reproductions are shown to be prompt copying, external retrieval, or coincidental short overlap rather than generation of memorized training passages.
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
Primary studies published before the post document repeated reproduction of training material. The GPT-2 study includes a 1,450-line code file and complete license texts. The later study finds many distinct verbatim sequences across several models. These results support the ordinary meaning of lengthy, repeated memorization without requiring a particular numerical threshold.
Limitations
- Extraction methods can involve deliberate prompting and specialized decoding.
- The evidence does not establish how often ordinary users encounter lengthy reproductions.
- It does not establish a universal training-data threshold or show that all training material is memorized.
Evidence (2 items)
Supports the claim: Carlini et al., arXiv,
1450 lines
Why it matters here: The GPT-2 study documents an entire code file and complete license texts among its extended extractions. These are multiple substantial reproductions, well beyond ordinary short-phrase overlap.
- Publisher
- Carlini et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Nasr et al., arXiv,
0.1% and 1%
Why it matters here: The study matches generated outputs against the open models' training datasets, requiring verbatim sequences of at least 50 tokens. It reports hundreds of thousands to millions of distinct matching sequences, independently supporting repeated memorization.
- Publisher
- Nasr et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument
LLMs don’t store things in simple databases, but when presented with enough data, their regeneration of distributed bits has demonstrably duplicated lengthy training materials multiple times, a phenomenon that is called memorization in the field.
- Published
- 2024-02-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deconstructing-geoffrey-hintons-weakest-530eb6-a1: 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-deconstructing-geoffrey-hintons-weakest-530eb6-a1.