“noncanonical reorderings such as these remain problematic for current AI approaches.)”
Our reading, the sentence that was checked:
Current AI approaches still struggle with noncanonical word orders such as 'the mat is on the cat'.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Three ideas from linguistics that everyone in AI should know, paragraph 37.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-three-ideas-from-linguistics-that-7845de-a2
We recorded 26 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 2022-06-22. 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
- noncanonical reorderings such as these
- remain problematic
- current AI approaches
- the mat is on the cat
Our interpretation
- The examples concern unusual semantic role or relation reversals, rather than necessarily ungrammatical syntax.
- Problematic means representative contemporary approaches show substantial, repeatable difficulty preserving the changed meaning. It does not require every model to fail every example.
- Because AI evaluations have no regular reporting quarter, current is assessed using the current calendar year through publication: 2022-01-01 to 2022-06-22.
Counts as supported if: Controlled evaluations of representative contemporary models show substantial failures distinguishing sentences or images whose meanings change when the same words exchange roles or positions.
Counts as contradicted if: Broad, representative contemporary evaluations show reliable handling of these semantic reversals, including unusual configurations.
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
The April 2022 Winoground evaluation tested image-caption matching with captions containing the same words in different orders [e1]. All tested models scored below random chance on the group measure, supporting substantial difficulty handling meaning-changing reorderings in contemporary vision-language models [e1]. A November 2022 analysis provides a later qualification: failures also involved combining visual and linguistic information, so they cannot be attributed solely to word-order processing [e2].
Limitations
- The evidence concerns tested vision-language models, not every contemporary AI approach [e1].
- The packet does not quantify performance on the exact cat-and-mat example.
- The November analysis falls outside the January 1–June 22, 2022 target period and serves only as methodological context [e2].
Evidence (2 items)
Supports the claim: Thrush et al., CVPR 2022 / arXiv,
All of the models are below random chance here as well.
Why it matters here: Winoground requires matching two images to captions containing the same words in different orders. Its poor group scores across contemporary vision-language models support substantial difficulty preserving the changed relationships.
- Publisher
- Thrush et al., CVPR 2022 / arXiv
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Diwan et al., arXiv,
meaningful model failures.
Why it matters here: This later analysis finds that visual discrimination and combining visual with linguistic information also affect performance. It limits any inference that word-order processing alone caused the observed failures.
- Publisher
- Diwan et al., arXiv
- 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, Three ideas from linguistics that everyone in AI should know
noncanonical reorderings such as these remain problematic for current AI approaches.)
- Where
- paragraph 37
- Published
- 2022-06-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-three-ideas-from-linguistics-that-7845de-a2: 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-three-ideas-from-linguistics-that-7845de-a2.