“But the variety of the examples along with the technical literature on memorization and data leakage in large language models make clear that this kind of reconstruction certainly can happen with some regularity.”
Our reading, the sentence that was checked:
Reconstruction of training text by large language models can happen with some regularity.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument, paragraph 52 (footnote 2).
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deconstructing-geoffrey-hintons-weakest-400537-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)
From the speaker
- reconstruction
- large language models
- certainly can happen with some regularity
- technical literature on memorization and data leakage
Our interpretation
- Reconstruction means generating substantial verbatim passages from training text, beyond merely reproducing text supplied in the prompt.
- The plain reading of 'can happen with some regularity' is a recurring, experimentally measurable phenomenon across examples. It does not specify a percentage of ordinary conversations or require every model to reproduce training text.
- The claim concerns an established capability as of 2024-02-05 and expressly invokes earlier research. The target period therefore ends on the post date and has no stated beginning. This capability has no quarterly or annual reporting period.
Counts as supported if: Pre-publication experiments demonstrate repeated verbatim reproduction of training passages across many examples, with counts or rates that establish recurrence rather than an isolated anecdote.
Counts as contradicted if: Reliable experiments show that the purported repeated reconstructions were not training-text reproductions, or that reconstruction was confined to isolated, unrepeatable cases under the tested conditions.
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
Research available before the post establishes recurring reconstruction. Nasr and colleagues reported extracting more than 10,000 distinct examples from ChatGPT under a targeted attack. Lee and colleagues separately measured verbatim training-text copying in more than 1% of unprompted output in their experiments. These findings support the ordinary reading that reconstruction can occur repeatedly. They do not establish a universal frequency, and the deduplication findings show that training practices matter.
Limitations
- Marcus supplies no numerical meaning for 'some regularity'.
- Experimental output rates do not directly measure ordinary chatbot use.
- The cited evidence supports recurrence, not universal susceptibility or any particular account of how model weights represent text.
Evidence (3 items)
Supports the claim: arXiv; Nasr and colleagues,
over 10,000 unique verbatim-memorized
Why it matters here: Thousands of distinct extracted examples establish recurrence under a targeted attack. Because ChatGPT's training corpus was unavailable, the researchers checked matches against a large auxiliary corpus; this is less direct than verification against a disclosed training set.
- Publisher
- arXiv; Nasr and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Lee and colleagues,
copied verbatim
Why it matters here: The measured share of unprompted output demonstrates recurring training-text reproduction beyond targeted chatbot attacks. The percentage describes generated output in these experiments, not the percentage of ordinary conversations affected.
- Publisher
- arXiv; Lee and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Kandpal, Wallace and Raffel,
more secure
Why it matters here: Deduplication reduces vulnerability, so extraction results cannot establish a uniform rate across models and training practices. This qualifies the scope of the supporting evidence without contradicting the claim that reconstruction can recur.
- Publisher
- arXiv; Kandpal, Wallace and Raffel
- 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
But the variety of the examples along with the technical literature on memorization and data leakage in large language models make clear that this kind of reconstruction certainly can happen with some regularity.
- Published
- 2024-02-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deconstructing-geoffrey-hintons-weakest-400537-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-400537-a1.