“LLMs do (sometimes) regurgitate large chunks of effectively stored text, and they don’t understand anything at a deep level.”
Our reading, the sentence that was checked:
LLMs sometimes regurgitate large chunks of effectively stored text and do not understand anything at a deep level.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument, paragraph 38.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deconstructing-geoffrey-hintons-weakest-cb96b8-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
- sometimes (qualifier, stated by the speaker)
From the speaker
- LLMs do (sometimes) regurgitate
- large chunks of effectively stored text
- they don’t understand anything at a deep level
Our interpretation
- This is a compound claim. Demonstrating memorization alone does not establish the assertion about understanding.
- The ordinary reading of deep understanding is a capacity to use learned concepts reliably in unfamiliar situations, including reasoning about causes and everyday situations. It does not require consciousness or perfect accuracy.
- The target is the latest cumulative public evidence about models existing by February 5, 2024. These capabilities have no standard reporting interval, and the post supplies no beginning date.
- Large chunks means extended passages whose reproduction cannot reasonably be explained by coincidental overlap in ordinary phrases. No numerical length threshold is attributed to Marcus.
Counts as supported if: Documented extraction establishes reproduction of extended training passages, and sufficiently broad tests establish the claimed absence of robust conceptual understanding in contemporary LLMs.
Counts as contradicted if: A contemporary LLM demonstrates robust conceptual understanding under a defensible test of the ordinary reading. Merely passing familiar benchmarks would not suffice. Failure to extract text in selected trials would not disprove the existential memorization assertion.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
Researchers demonstrated extended verbatim reproduction, supporting the first assertion. The second assertion is much broader. Generalization failures establish limitations, while representation research cautions against equating prediction with surface matching. Neither establishes whether contemporary LLMs understood nothing at a deep level. The compound claim therefore remains unresolved; partial support alone is not a mixed outcome.
Limitations
- Closed-model training membership is inferred from matching pre-existing material rather than inspection of the complete private training corpus.
- The understanding studies cover limited tasks and model families.
- The Othello result is contextual counterevidence from a synthetic setting.
Evidence (3 items)
Supports the claim: Nasr, Carlini and collaborators,
matches 100% word-for-word
Why it matters here: The researchers display extended verbatim outputs and describe checking them against pre-existing web data. This supports the memorization component, without establishing anything about the absence of understanding.
- Publisher
- Nasr, Carlini and collaborators
- Kind of source
- Demonstration
- Describes events of
- Public from
Context only: Berglund et al., arXiv,
a surprising failure of generalization
Why it matters here: Controlled experiments found failures to infer reversed relationships after training on the forward relationship. This establishes a specific limitation, not the absence of deep understanding in every domain.
- Publisher
- Berglund et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Li et al., arXiv,
an emergent world representation.
Why it matters here: The study reports an internal board representation and interventions affecting predictions. This challenges a general inference from sequence prediction to mere surface matching, but the synthetic Othello model does not establish deep understanding in general-purpose LLMs.
- Publisher
- Li et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A defensible operational test of deep understanding, with broad controlled evaluations of models available by February 5, 2024. Memorization is already documented.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument
LLMs do (sometimes) regurgitate large chunks of effectively stored text, and they don’t understand anything at a deep level.
- Where
- paragraph 38
- Published
- 2024-02-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deconstructing-geoffrey-hintons-weakest-cb96b8-a1: 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-deconstructing-geoffrey-hintons-weakest-cb96b8-a1.