“And what Southen and I showed is that generative image programs are prone to exactly the same kind of regurgitation that is at issue.”
Our reading, the sentence that was checked:
Generative image programs are prone to the same kind of regurgitation of copyrighted material as text models.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The desperate race to save Generative AI, paragraph 31.
- 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-e83129-a1
We recorded 13 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)
From the speaker
- generative image programs
- are prone to
- exactly the same kind of regurgitation
Our interpretation
- The plain reading is that image generators have a demonstrated susceptibility to reproducing existing source images, comparable to text models reproducing source passages.
- Near-copies must preserve identifiable source content; similarity of style alone does not qualify.
- The comparison concerns the kind of failure, not equal reproduction rates across image and text models.
- The observation concerns capabilities demonstrated by the post date. These systems have no regular reporting period, and the retrospective wording supplies no starting date, so the target period has an unspecified start.
- This assesses the technical phenomenon, not whether a particular output legally infringes copyright or whether Marcus and Southen first discovered it.
Counts as supported if: Independent experiments on image generators available before the post document reproduction of identifiable training images or substantial source content, using source comparisons and described methods.
Counts as contradicted if: Independent examination shows that the purported reproductions are merely stylistic similarities, supplied input images, or retrieval results, and credible tests fail to substantiate generation of identifiable source content.
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
Independent research available before the post showed image generators memorizing and reproducing training images. A separate study identified copying through comparisons with source images. This supports the ordinary technical reading of regurgitation. The evidence does not establish that copying dominates normal use, occurs equally often in image and text models, or necessarily constitutes copyright infringement.
Limitations
- The result depends on reading prone as demonstrated susceptibility rather than a specified high frequency.
- Model versions, prompts, and extraction methods affect the findings.
- The attribution to Marcus and Southen is not independently adjudicated.
Evidence (3 items)
Supports the claim: arXiv; Nicholas Carlini and colleagues,
memorize individual images
Why it matters here: The experiment establishes generation of memorized training images. This directly supports the technical meaning of regurgitation, without establishing infringement or a typical-use frequency.
- Publisher
- arXiv; Nicholas Carlini and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Gowthami Somepalli and colleagues,
copy from their training data
Why it matters here: A separate research team compared generated images with training samples and identified copying in Stable Diffusion. This corroborates the phenomenon independently of Marcus and Southen.
- Publisher
- arXiv; Gowthami Somepalli and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Ryan Webster,
typically is only possible for very few samples.
Why it matters here: This limits a stronger reading that image generators routinely reproduce arbitrary training images. It does not negate demonstrated susceptibility.
- Publisher
- arXiv; Ryan Webster
- 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, The desperate race to save Generative AI
And what Southen and I showed is that generative image programs are prone to exactly the same kind of regurgitation that is at issue.
- Where
- paragraph 31
- Published
- 2024-01-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-desperate-race-to-save-generative-e83129-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-the-desperate-race-to-save-generative-e83129-a1.