“As Ernie Davis, Scott Aaronson, and I showed, a week or so after the release, it’s easily fooled, especially with complex sentences and words like “similar” that would require a deep understanding of language to be interpreted correctly”
Our reading, the sentence that was checked:
DALL-E 2 is easily fooled by complex sentences and words requiring deep language understanding.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Horse rides Astronaut, redux, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-horse-rides-astronaut-redux-24077b-a1
We recorded 8 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
- 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
- especially complex sentences and words like 'similar' (qualifier, stated by the speaker)
From the speaker
- a week or so after the release
- it’s easily fooled
- especially with complex sentences and words like “similar”
- a red basketball with flowers on it, in front of blue one with with a similar pattern
Our interpretation
- This passage quotes the May 2022 essay and describes early DALL-E 2. The target period is April–May 2022, rather than the model's condition in February 2024.
- The ordinary behavioral reading is that complex compositional instructions readily produce failures to depict the requested objects, attributes and relations.
- No numerical failure threshold is supplied. I require a repeated failure pattern, rather than one unsuccessful image.
- The reference to deep understanding is not treated as a separately established explanation of the model's internal processing.
Counts as supported if: Admissible prompt-and-output evidence about early DALL-E 2 demonstrates repeated failures on complex sentences and relational expressions of the kind described.
Counts as contradicted if: Comparable evidence shows reliable performance on those instructions and materially undermines the reported failure pattern. Occasional successes alone do not contradict the claim.
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 evidence supports a narrower early limitation: DALL-E 2 could misassign attributes to objects. A later independent study also found poor relational performance. Neither retrieved source settles the full claim about complex sentences and expressions such as similar in April–May 2022. The directly relevant named studies cannot establish the outcome because Marcus coauthored them. This is an evidence gap, not evidence that his observation was false.
Limitations
- A broader reading covering all of 2022 would admit more independent evidence.
- The assessment does not establish or refute a theory of deep language understanding.
Evidence (2 items)
Supports the claim: OpenAI researchers; arXiv,
worse at binding attributes to objects
Why it matters here: The model authors document attribute-assignment failures, illustrated with colored cubes. This supports a component of Marcus's observation, but does not establish the full complex-sentence or similar-prompt claim.
- Publisher
- OpenAI researchers; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Colin Conwell and Tomer Ullman; arXiv,
only ~22% of images matched basic relation prompts.
Why it matters here: An independent study reports poor relational performance. The retrieved record does not date its tests to April–May, so it cannot settle this historical claim.
- Where
- Abstract
- Publisher
- Colin Conwell and Tomer Ullman; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Independent dated output records from early DALL-E 2, with complete batches for the cited prompt and comparable complex instructions.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Horse rides Astronaut, redux
As Ernie Davis, Scott Aaronson, and I showed, a week or so after the release, it’s easily fooled, especially with complex sentences and words like “similar” that would require a deep understanding of language to be interpreted correctly
- Where
- paragraph 5
- Published
- 2024-02-13
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-horse-rides-astronaut-redux-24077b-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-24077b-a1.