“Of course we see the same thing over and over in language systems, too.”
Our reading, the sentence that was checked:
Language systems repeatedly show the same mix of fluent output and lack of understanding.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, DeepFakes versus Deep Understanding, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deepfakes-versus-deep-understanding-b2d26e-a1
We recorded 4 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 2022-10-02. 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
- the same thing
- over and over
- in language systems, too
- what should I do if my pig starts flying?
Our interpretation
- The ordinary reading is that contemporary language systems repeatedly produce competent-looking language while making basic commonsense mistakes or accepting false premises.
- This does not require proving that models possess no understanding whatsoever, or that every response fails.
- With no reporting cycle or exact period specified, use the current calendar year through publication, January 1 through October 2, 2022, for the ongoing observation.
- Systematic evaluations and documented recurring failure modes can establish repetition without reproducing the exact flying-pig responses.
Counts as supported if: Contemporary primary reports or reproducible evaluations document recurring commonsense or false-premise failures alongside competent language performance.
Counts as contradicted if: Reliable contemporary evaluations establish that the alleged recurring pattern is an artifact or misdescription and that the relevant systems reliably handle the ordinary commonsense and false-premise cases at issue.
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 behavioral claim is supported. The InstructGPT report documents recurring simple mistakes, including accepting false premises, alongside competent language performance. An independent 2022 study finds systematic commonsense weaknesses. Research showing that input presentation improves performance limits a stronger philosophical conclusion: these records establish unreliable performance, not the absence of all understanding.
Limitations
- Support concerns the recurring behavioral pattern, not the exact anecdotes.
- The studies cover particular models, prompts, and tasks.
- The July source supplies only a publication month; July 31 is a conservative availability bound.
Evidence (3 items)
Supports the claim: arXiv; OpenAI research authors,
InstructGPT still makes simple mistakes.
Why it matters here: The surrounding results describe recurrent false-premise acceptance, excessive hedging, and constraint failures despite strong language-task performance. These directly match the behavioral pattern Marcus describes.
- Publisher
- arXiv; OpenAI research authors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Association for Computational Linguistics,
GPT-3’s performance was mostly at chance
Why it matters here: The study systematically tests commonsense implications and presuppositions. Its recurring weaknesses corroborate the commonsense component, although it measures sentence probabilities rather than fluent conversational answers.
- Where
- Abstract
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Association for Computational Linguistics,
outperform
Why it matters here: The authors improve commonsense generation by changing the order of input concepts and argue that existing capability is underestimated. This qualifies any inference from failures to a complete absence of understanding, without refuting recurring failures.
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, DeepFakes versus Deep Understanding
Of course we see the same thing over and over in language systems, too.
- Where
- paragraph 5
- Published
- 2022-10-02
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deepfakes-versus-deep-understanding-b2d26e-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-deepfakes-versus-deep-understanding-b2d26e-a1.