“The thing is, current AI systems that (roughly speaking) do a lot of cutting and pasting without those grammar rules and trees aren’t really all that close to understanding human language.”
Our reading, the sentence that was checked:
Current AI systems that, roughly speaking, do a lot of cutting and pasting without grammar rules and trees aren't really all that close to understanding human language.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Learning Language is Harder Than You Think, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-learning-language-is-harder-than-b076d7-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 2022-07-20. 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
- systems that, roughly speaking, do a lot of cutting and pasting without grammar rules and trees (scope, stated by the speaker)
- aren't really all that close (qualifier, stated by the speaker)
From the speaker
- current AI systems
- roughly speaking
- without those grammar rules and trees
- aren’t really all that close to understanding human language
Our interpretation
- The ordinary reading is that contemporary neural language systems remained substantially below reliable human language comprehension across varied situations.
- The cutting-and-pasting description identifies the systems under discussion; it is not treated as a separate, literal claim about their implementation.
- Current refers to systems existing on July 20, 2022. With no regular reporting cycle for capabilities, I use current-year evaluations through that date as the observation window.
- Benchmark performance is evidence about particular abilities, not automatically a measure of distance from general language understanding.
Counts as supported if: Broad, representative evaluations establish substantial and persistent deficits relative to human comprehension across syntax, semantics, and contextual interpretation.
Counts as contradicted if: Representative systems in the described class demonstrate reliably human-comparable comprehension across those areas and varied unfamiliar situations.
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
Contemporary evaluations document substantial weaknesses and strong successes. BIG-bench found large gaps for its tested models, while PaLM exceeded average human performance on selected tasks. Neither finding settles the broader claim that systems were not close to understanding language. Turning these results into that verdict would require a threshold Marcus did not provide.
Limitations
- Benchmark achievement and general language comprehension are different claims.
- The evidence concerns capabilities documented during the observation window; exact experiment dates are not supplied.
Evidence (2 items)
Context only: arXiv; Srivastava et al.,
strongest models perform poorly overall
Why it matters here: BIG-bench documents substantial gaps for its evaluated models. Its deliberately challenging task collection does not itself define how close a system is to general language understanding.
- Publisher
- arXiv; Srivastava et al.
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Chowdhery et al.,
average human performance
Why it matters here: PaLM's human-comparable or better results on selected tasks are material counterevidence to a uniformly poor picture. They do not establish reliable comprehension across ordinary language situations.
- Publisher
- arXiv; Chowdhery et al.
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A justified assessment of July 2022 systems across representative comprehension tasks, with human comparisons and a defensible criterion for being close to general competence.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Learning Language is Harder Than You Think
The thing is, current AI systems that (roughly speaking) do a lot of cutting and pasting without those grammar rules and trees aren’t really all that close to understanding human language.
- Where
- paragraph 10
- Published
- 2022-07-20
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-learning-language-is-harder-than-b076d7-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-learning-language-is-harder-than-b076d7-a1.