“And they have such a superficial understanding of their own output that they can’t begin to fact check it.”
Our reading, the sentence that was checked:
LLMs have such a superficial understanding of their own output that they can't even begin to fact-check it.
Assessment: Contradicted Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why DO large language models hallucinate?, paragraph 27.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-why-do-large-language-models-hallucinate-fc26ca-a1
We recorded 9 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-05-05. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- such a superficial understanding of their own output
- that they can’t begin to fact check it
Our interpretation
- The ordinary reading of can’t begin is absence of meaningful self-fact-checking ability, rather than merely imperfect reliability.
- Self-fact-checking means identifying the truth or falsity of factual claims generated by the same model, assessed against independent human judgments.
- Prompting a model to assess its generated claims qualifies. A separate search engine or trained detector is an additional component and must be identified.
- The alleged superficial understanding is a causal explanation, not something established simply by observing errors.
- Use the current year's pre-publication reporting window, 2025-01-01 through 2025-05-05.
Counts as supported if: Controlled evaluations show no meaningful ability to distinguish correct from incorrect self-generated factual claims, and evidence connects that inability to the alleged representational deficit.
Counts as contradicted if: Controlled evaluations demonstrate meaningful identification of errors in the same model’s factual outputs. Such evidence contradicts the asserted inability without proving comprehensive understanding or dependable autonomous verification.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Contradicted
The dated evidence goes against 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 direct self-evaluation study shows models identifying substantial numbers of false claims in their own generated reports without retrieval. That is meaningful fact-checking and contradicts can’t begin on its ordinary reading. The same study supplies an important qualification: many claims received no assessment, and some judgments were wrong. Thus it does not establish reliable autonomous fact-checking. The proposed explanation involving superficial understanding remains unproven.
Limitations
- The verdict turns on the stronger inability claim. A claim of unreliable self-checking would receive a different assessment.
- Evidence item e1 also contains the main counterweight: substantial abstention and incorrect judgments.
- Successful error identification does not establish human-like understanding.
- The reported experimental dates are less precise than the publication date.
Evidence (2 items)
Goes against the claim: Jiayi Yao, Haibo Sun and Nianwen Xue / arXiv,
correct and incorrect assessments
Why it matters here: In the no-retrieval self-evaluation setting, Table 5 reports correct assessments for 201 of 532 false GPT-generated claims and 229 of 537 false GLM-generated claims. Substantial abstention remains, but identifying these errors contradicts the claimed absence of meaningful self-fact-checking ability.
- Publisher
- Jiayi Yao, Haibo Sun and Nianwen Xue / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: YiQiu Guo and colleagues / arXiv,
reducing factual hallucinations
Why it matters here: DSVD reports improved factuality using internal-state probes and rollback. Because the detector receives additional training using known answers, this is context about available internal signals rather than decisive evidence of an unaided model’s self-checking ability.
- Publisher
- YiQiu Guo and colleagues / 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, Why DO large language models hallucinate?
And they have such a superficial understanding of their own output that they can’t begin to fact check it.
- Where
- paragraph 27
- Published
- 2025-05-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-why-do-large-language-models-hallucinate-fc26ca-a1: Current. Outcome: Contradicted. 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-why-do-large-language-models-hallucinate-fc26ca-a1.