“It’s likewise a flaw that that current AI cannot learn without a tendency to inadvertently regurgitate big bits of text and images that look uncannily like their sources, but not a logical necessity.”
Our reading, the sentence that was checked:
Current AI cannot learn without a tendency to inadvertently regurgitate big bits of text and images that look uncannily like their sources, which is a flaw but not a logical necessity.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The desperate race to save Generative AI, paragraph 17.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-the-desperate-race-to-save-generative-4d75f5-a2
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 2024-01-08. 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
- current AI (scope, stated by the speaker)
From the speaker
- current AI cannot learn without a tendency
- inadvertently regurgitate big bits of text and images
- look uncannily like their sources
- not a logical necessity
Our interpretation
- Current AI refers to contemporary text and image generators in the surrounding discussion.
- Substantial reproduction means identifiable passages or near-duplicate images, rather than short common phrases or general stylistic resemblance.
- The plain reading includes both an observed copying tendency and the stronger assertion that contemporary learning cannot avoid that tendency.
- Deliberate extraction establishes susceptibility but does not measure accidental reproduction during ordinary use.
- The reporting window is the latest completed calendar research year and elapsed days before publication: 2023-01-01 through 2024-01-08.
Counts as supported if: Research demonstrates substantial source reproduction in contemporary text and image generators and establishes that the tendency cannot be avoided while retaining useful learning under contemporary methods.
Counts as contradicted if: A contemporary learning method demonstrably avoids the relevant copying tendency while retaining useful learning. Merely reducing leakage or failing to find copies is insufficient.
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
The January 30, 2023 study extracted training images from tested diffusion models [e1]. The November 28, 2023 study documented long memorized passages extracted from ChatGPT through an attack [e2]. These findings establish susceptibility in the tested systems, but do not establish that useful contemporary learning cannot avoid it or measure inadvertent reproduction during ordinary use [e1, e2]. The privacy-preserving method and OpenAI's January 8, 2024 mitigation statement do not demonstrate elimination of substantial reproduction either [e3, e4]. The full claim therefore remains unresolved [e1–e4].
Limitations
- The extraction studies do not establish an unavoidable tendency across contemporary text and image generators [e1, e2].
- Differential privacy does not guarantee that outputs never resemble training material [e3].
- OpenAI's statement supplies no independent measurement of elimination; its publication order relative to Marcus's post on the same day is unknown [e4].
Evidence (4 items)
Supports the claim: arXiv; Nicholas Carlini and coauthors,
emit them at generation
Why it matters here: The study demonstrates training-image memorization and reproduction. It supports the observed tendency, but does not prove that every contemporary learning method must retain it.
- Publisher
- arXiv; Nicholas Carlini and coauthors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Milad Nasr and coauthors,
memorized training examples
Why it matters here: The paper documents substantial text reproduction from ChatGPT using an extraction attack. This establishes susceptibility, not the frequency of accidental copying in ordinary use or its inevitability.
- Publisher
- arXiv; Milad Nasr and coauthors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Tim Dockhorn and coauthors,
enforce DP guarantees.
Why it matters here: This pre-existing method demonstrates useful image learning with formal privacy protections and bears against an unrestricted inevitability claim. Differential privacy does not guarantee that no generated image will ever resemble training material.
- Publisher
- arXiv; Tim Dockhorn and coauthors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: OpenAI,
limit inadvertent memorization
Why it matters here: OpenAI describes mitigation and disputes whether induced examples represent normal use. Its statement supplies no independent measurement establishing elimination of the risk.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
A precise definition of the claimed tendency, followed by evidence that useful learning necessarily retains it under the relevant contemporary methods, or a method available in the target period demonstrably avoiding it while retaining useful learning. Reduced leakage or unsuccessful extraction alone would not settle the claim.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, The desperate race to save Generative AI
It’s likewise a flaw that that current AI cannot learn without a tendency to inadvertently regurgitate big bits of text and images that look uncannily like their sources, but not a logical necessity.
- Where
- paragraph 17
- Published
- 2024-01-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-desperate-race-to-save-generative-4d75f5-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-the-desperate-race-to-save-generative-4d75f5-a2.