“In other words, the problem was with its language understanding, rather than with illustration per se. The model was perfectly capable of drawing a horse riding an astronaut, it just didn’t know that the phrase horse rides astronaut was a request for a horse riding an astronaut.”
Our reading, the sentence that was checked:
DALL-E's failure on 'horse rides astronaut' was due to deficient language understanding, not illustration ability.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Horse rides Astronaut, redux, paragraph 17.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-horse-rides-astronaut-redux-a310b4-a1
We recorded 5 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
- the problem was with its language understanding
- rather than with illustration per se
- perfectly capable of drawing a horse riding an astronaut
- back then, in May 2022
Our interpretation
- DALL-E means DALL-E 2 in the May 2022 episode, not Imagen or a later model.
- The ordinary reading is that the model could render the requested scene but failed to map the original wording to that scene.
- Successful rephrasing would establish rendering capability and prompt sensitivity. Attributing the failure specifically to language understanding also requires evidence distinguishing that explanation from sampling variation or other parts of image generation.
Counts as supported if: Records from the same DALL-E 2 version show the requested scene rendered successfully with equivalent rephrasing, repeated failures with the original wording, and evidence that the difference reflects interpretation of the prompt.
Counts as contradicted if: Controlled evidence shows that the original wording was interpreted correctly and the failure arose elsewhere, or establishes that the claimed successful rendering was not produced by the relevant DALL-E 2 model.
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
The sources establish related limitations and a reported failed attempt. They do not establish the full causal claim about DALL-E 2. The successful rephrasing example also has a model-identity problem: the earlier account attributes it to Imagen. Even a verified success after rephrasing would demonstrate rendering capability without necessarily isolating language understanding as the cause.
Limitations
- The original images and complete sampling records were not available in the retrieved text.
- The technical paper concerns related failures, not this particular experiment.
Evidence (2 items)
Context only: OpenAI researchers via arXiv,
binding attributes to objects
Why it matters here: The authors document related compositional limitations. This earlier result does not establish the cause of the particular May failure or successful rendering of the requested scene.
- Publisher
- OpenAI researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Joscha Bach via Twitter,
it did not work
Why it matters here: This is a contemporaneous report of an unsuccessful attempt. The remark about weight is a joke, not evidence of a causal mechanism.
- Publisher
- Joscha Bach via Twitter
- Kind of source
- Demonstration
- Describes events of
- Public from
What would settle it
Authenticated paired experiments using the May 2022 DALL-E 2 version, with repeated trials and evidence distinguishing prompt interpretation from other generation failures.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Horse rides Astronaut, redux
In other words, the problem was with its language understanding, rather than with illustration per se. The model was perfectly capable of drawing a horse riding an astronaut, it just didn’t know that the phrase horse rides astronaut was a request for a horse riding an astronaut.
- Where
- paragraph 17
- Published
- 2024-02-13
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-horse-rides-astronaut-redux-a310b4-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-horse-rides-astronaut-redux-a310b4-a1.