“That’s why statistically improbable requests like horses riding astronauts have always been challenging for generative image systems.”
Our reading, the sentence that was checked:
Because they rely on statistics instead of understanding, generative image systems have always struggled with statistically improbable requests such as horses riding astronauts.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Statistics versus Understanding: The Essence of What Ails Generative AI, paragraph 7.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-statistics-versus-understanding-the-990d7f-a1
We recorded 6 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 2024-02-13. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- That’s why
- statistically improbable requests
- horses riding astronauts
- have always been challenging
- generative image systems
Our interpretation
- The ordinary reading is recurring difficulty with unusual compositions among image generators available before publication, not that every attempt necessarily fails.
- The causal explanation is that learned preferences for familiar relationships interfere with following an unfamiliar relationship specified in the prompt.
- Historical difficulty and its proposed cause must both be assessed. Documenting errors alone does not settle the causal explanation.
- The historical scope ends on February 13, 2024; the source supplies no beginning date.
Counts as supported if: Representative historical evaluations show recurring difficulty with rare compositions, and controlled comparisons establish that learned frequency preferences materially cause those errors.
Counts as contradicted if: Representative historical evaluations show reliable performance on the relevant rare compositions, or controlled causal tests show that the proposed frequency mechanism does not explain the observed difficulty.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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
There is relevant support: a 2023 study reports that training frequency predicts Stable Diffusion 2.1's success and difficulty with reversed relationships. This is stronger evidence than selected failed images. However, the evidence retrieved does not settle Marcus's broader causal explanation across image systems or establish the proposed contrast between statistics and understanding. The narrower frequency-related finding is supported; the complete causal claim remains unresolved.
Limitations
- The most directly relevant study concerns one model.
- The full PDF could not be read with the required tool.
- Occasional successful generations would not, by themselves, contradict a claim of difficulty.
Evidence (1 item)
Supports the claim: Charles Lovering and Ellie Pavlick, arXiv,
the flipped triad.
Why it matters here: The study reports that Stable Diffusion 2.1's performance tracks training frequency and worsens on reversed relationships as the familiar relationship becomes more frequent. This supports the proposed frequency mechanism in one historical system, but does not establish the full causal generalization.
- Publisher
- Charles Lovering and Ellie Pavlick, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled evaluations of representative pre-publication models that vary relationship frequency while holding competing explanations constant, alongside repeated tests of rare and familiar compositions.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Statistics versus Understanding: The Essence of What Ails Generative AI
That’s why statistically improbable requests like horses riding astronauts have always been challenging for generative image systems.
- Where
- paragraph 7
- Published
- 2024-02-13
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-statistics-versus-understanding-the-990d7f-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-statistics-versus-understanding-the-990d7f-a1.