“The dirty secret in current AI is that for all the bloviating about how current systems have mastered human language, they are still really weak on compositionality, which most linguists (since Frege, over a century ago) would agree is at the very core of how language works.”
Our reading, the sentence that was checked:
Current AI systems are still very weak at compositionality, which is central to how language works.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?, paragraph 13.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-did-googleai-just-snooker-one-of-6bfe53-a1
We recorded 12 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 2022-09-14. 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
- current AI
- still really weak on compositionality
- at the very core of how language works
Our interpretation
- The plain reading is that prominent contemporary neural language and vision-language systems had substantial weaknesses in combining familiar elements according to sentence structure.
- It does not assert that every AI architecture fails every compositional task.
- Because these systems have no common quarterly reporting period, use the 2022 year-to-post evaluation period. The context explicitly discusses contemporary 2022 models and evaluations.
- Treat compositionality's central role as an established linguistic principle, rather than infer a numerical majority of linguists without a survey.
Counts as supported if: Contemporary evaluations show substantial structural or compositional generalization failures across prominent model families, while linguistic research recognizes compositionality as a basic property of language.
Counts as contradicted if: Representative contemporary systems reliably handle unfamiliar compositional structures across relevant tasks, making the characterization of a substantial general weakness untenable.
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
Contemporary research supports substantial compositional weaknesses in prominent neural systems. Winoground found large human-model gaps, while COGS evaluations found poor structural generalization in sequence-to-sequence models. The qualification matters: specialized compositional parsers performed strongly, and lexical generalization was much easier. The evidence supports the contextual claim about a substantial weakness, not a universal inability across all AI systems.
Limitations
- The result depends on reading current AI as prominent contemporary neural systems rather than every AI architecture.
- Compositional performance varies substantially by task.
- The linguistic background supports the principle's importance but does not measure majority opinion.
Evidence (3 items)
Supports the claim: arXiv; Thrush and colleagues,
models struggled
Why it matters here: Winoground found large human-model gaps across several vision-language model families on image-caption matching that depends on composition. The finding supports a broad weakness within the tested contemporary systems.
- Publisher
- arXiv; Thrush and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Weißenhorn and colleagues,
low accuracies
Why it matters here: The COGS comparison supports a structural-generalization weakness in sequence-to-sequence models, including BART. Its successful compositional parsers and strong lexical results are material counterevidence to any universal claim that all AI systems lack compositional ability.
- Publisher
- arXiv; Weißenhorn and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Kim and Linzen,
Natural language is characterized by compositionality
Why it matters here: This research states the linguistic principle underlying the claim. Its older experimental findings are not used as outcome evidence for 2022 systems.
- Publisher
- arXiv; Kim and Linzen
- 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, Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
The dirty secret in current AI is that for all the bloviating about how current systems have mastered human language, they are still really weak on compositionality, which most linguists (since Frege, over a century ago) would agree is at the very core of how language works.
- Where
- paragraph 13
- Published
- 2022-09-14
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-did-googleai-just-snooker-one-of-6bfe53-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-did-googleai-just-snooker-one-of-6bfe53-a1.