“Large language models still make stuff up (“hallucinate”), still can’t be trusted in their reasoning, and remain unreliable.”
Our reading, the sentence that was checked:
LLMs still hallucinate and their reasoning cannot be trusted.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Six months after the Pause letter, AI is still erratic — and still something we should still be deeply concerned about, paragraph 12.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-six-months-after-the-pause-letter-aec4cd-a1
We recorded 9 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-09-21. 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
- Large language models still make stuff up
- hallucinate
- still can’t be trusted in their reasoning
- remain unreliable
Our interpretation
- The ordinary reading is that contemporary general-purpose LLMs still produce unsupported factual assertions and material reasoning errors, so unchecked answers are not dependable.
- This does not assert that every answer is wrong or that no bounded application can use an LLM reliably.
- Factuality and reasoning require separate evidence; neither is inferred solely from the other.
- No measurement period is specified. I use the six-month retrospective named in the title, ending on publication, as the reporting window for the continuing state.
Counts as supported if: Contemporaneous evaluations document both unsupported factual output and substantial reasoning failures in leading LLMs.
Counts as contradicted if: Representative evaluations establish dependable factual accuracy and reasoning across the claimed general-purpose scope, with the alleged recurring failures absent.
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 evidence documents both parts of the claim. FActScore found substantial unsupported factual content in generated biographies. A separate planning study found frequent failures by GPT-4 on tasks with mechanically checkable solutions. This supports the ordinary reading that factual and reasoning outputs still needed verification. The planning study also found improvements and useful roles for external assistance; those findings do not establish that the underlying reliability problems had been solved.
Limitations
- The results concern specific models, tasks, and prompts rather than every possible use.
- The studies were reported in May; they are not a direct test of every model update available on September 21.
- The sources do not identify exact dates for every experimental run.
Evidence (2 items)
Supports the claim: Sewon Min and colleagues, via arXiv,
42.5% and 58.3%
Why it matters here: Human fact-checking found substantial unsupported content in generated biographies. These are factual-precision measurements against the study's knowledge source, not general error rates for every ChatGPT answer.
- Publisher
- Sewon Min and colleagues, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Karthik Valmeekam and colleagues, via arXiv,
approximately 34% in the Blocksworld dataset.
Why it matters here: Executable-plan checks expose substantial failures in a defined reasoning task. The paper also finds improvements over older models and benefits from external planners and verification, limiting any inference that LLM-assisted reasoning is invariably unreliable.
- Publisher
- Karthik Valmeekam and colleagues, via 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, Six months after the Pause letter, AI is still erratic — and still something we should still be deeply concerned about
Large language models still make stuff up (“hallucinate”), still can’t be trusted in their reasoning, and remain unreliable.
- Where
- paragraph 12
- Published
- 2023-09-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-six-months-after-the-pause-letter-aec4cd-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-six-months-after-the-pause-letter-aec4cd-a1.