“LLMs can still cannot be trusted:”
Our reading, the sentence that was checked:
LLMs still cannot be trusted.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Let’s be honest, Generative AI isn’t going all that well, paragraph 2.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-lets-be-honest-generative-ai-isnt-c5b210-a2
We recorded 7 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 2026-01-12. 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
- LLMs
- still
- cannot be trusted
Our interpretation
- The plain reading is that factual answers from leading general-purpose LLMs still require independent verification when accuracy matters.
- This does not mean that every answer is wrong or that no bounded application can be reliable.
- Because the statement describes an ongoing state without dates, use the latest completed quarter before publication, October–December 2025.
Counts as supported if: Contemporaneous evaluations show recurring factual errors and failures to acknowledge uncertainty across leading models and multiple domains.
Counts as contradicted if: Broad, representative evaluations show that leading models reliably answer or abstain across factual domains, making independent verification generally unnecessary.
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
In November 2025, AA-Omniscience reported persistent factuality and calibration weaknesses across frontier models on 6,000 questions spanning six domains [e1]. This supports the reading that factual answers still need independent verification when accuracy matters [e1]. In December 2025, OpenAI reported a hallucination rate below 1% for GPT-5.2 Thinking with browsing enabled across five domains in its production-conversation evaluation [e2]. That result supports reliable performance in the evaluated setting, without establishing that verification is generally unnecessary [e2].
Limitations
- The supplied AA-Omniscience abstract does not specify whether browsing was enabled [e1].
- The favorable browsing result is company-reported and uses LLM-based grading [e2].
- These evaluations do not establish a single error rate for ordinary LLM use, an acceptable risk threshold for every application, or unreliability in every use [e1, e2].
Evidence (2 items)
Supports the claim: Artificial Analysis / arXiv,
These results reveal persistent factuality and calibration weaknesses across frontier models.
Why it matters here: The evaluation covers 6,000 questions across six domains. Only three evaluated models scored above an index value representing equal numbers of correct and incorrect answers, supporting continued verification of factual outputs.
- Where
- Abstract
- Publisher
- Artificial Analysis / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: OpenAI,
<1% hallucination rate
Why it matters here: This materially limits any suggestion that factual errors are frequent in every setting. It measures a different workload with browsing and does not establish that general factual answers can safely go unchecked.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Let’s be honest, Generative AI isn’t going all that well
LLMs can still cannot be trusted:
- Where
- paragraph 2
- Published
- 2026-01-12
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-lets-be-honest-generative-ai-isnt-c5b210-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-lets-be-honest-generative-ai-isnt-c5b210-a2.