“Neural networks do effectively memorize some of what they are trained on, and indeed there is a whole subfield now about the topic.”
Our reading, the sentence that was checked:
Neural networks effectively memorize some of what they are trained on, and there is now a whole subfield devoted to this topic.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Deconstructing Geoffrey Hinton’s weakest argument, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deconstructing-geoffrey-hintons-weakest-d3a2f9-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
- some of what they are trained on (scope, stated by the speaker)
From the speaker
- Neural networks do effectively memorize
- some of what they are trained on
- a whole subfield now about the topic
Our interpretation
- Effective memorization means retaining training-specific information that can be recovered in model outputs; it does not require literal storage of documents.
- Some is an existence claim, not a claim that every network memorizes everything.
- Subfield means an established research line with multiple studies, methods, and follow-up questions, not formal institutional recognition.
- For this ongoing research claim, use the latest complete calendar year and current year-to-date, 2023 through the post date. Research has no standard quarterly reporting period; earlier studies supply historical context.
Counts as supported if: Experiments demonstrate recovery of training text from neural models, and multiple research studies investigate memorization or extraction.
Counts as contradicted if: The purported recoveries do not demonstrate retained training information, or the alleged research line is unsupported by identifiable studies.
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
Original experiments show that neural language models can reproduce training material, including exact passages. The extraction and quantitative memorization papers also establish a continuing research literature. This supports Marcus's limited claim about some training information; it does not imply that networks literally store all documents or that memorization explains every output.
Limitations
- For closed models, the researchers used an auxiliary corpus because the complete training data were unavailable.
- The demonstrated ChatGPT attack was specific to the tested model and method.
- Calling the research line a subfield is an ordinary descriptive judgment.
Evidence (2 items)
Supports the claim: Nasr and colleagues, arXiv,
verbatim-memorized training examples.
Why it matters here: The paper reports substantial extraction from neural language models and directly checks known training corpora for open models. Its methods and review of preceding studies also support the existence of an established research line.
- Where
- Section 5.3, Main Experimental Results; sections 2–3 for earlier work and open-model verification
- Publisher
- Nasr and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Carlini and colleagues, arXiv,
memorized training data.
Why it matters here: This study, initially submitted in 2022 and revised in 2023, investigates how memorization varies with model size, repetition, and prompting. It supplies historical evidence of a research program predating the post.
- Publisher
- Carlini and colleagues, 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
Neural networks do effectively memorize some of what they are trained on, and indeed there is a whole subfield now about the topic.
- Where
- paragraph 20
- Published
- 2024-02-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deconstructing-geoffrey-hintons-weakest-d3a2f9-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-d3a2f9-a1.