“When I think about how poor his company (Meta) is even at making LLMs that don’t hallucinate, I get worried.”
Our reading, the sentence that was checked:
Meta is poor at making LLMs that do not hallucinate.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, 𝐩(𝐝𝐨𝐨𝐦), paragraph 23.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-d28-267622-a1
We recorded 24 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 2023-08-27. 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
- his company (Meta)
- how poor
- making LLMs that don’t hallucinate
Our interpretation
- The ordinary reading is that Meta’s available LLMs remained materially unreliable at producing factual answers.
- The claim does not expressly say Meta was worse than competing developers.
- The target is the latest public model generation and its evaluations before the post: the July–August 2023 release period.
- Hallucination means fabricated or factually unsupported output. A combined truthfulness-and-informativeness score is relevant evidence, but its complement is not a hallucination rate.
Counts as supported if: Evaluations of Meta’s then-current LLMs show substantial factual reliability limitations, corroborated by documented hallucination behavior or explicit technical limitations.
Counts as contradicted if: Broad evaluations of those models demonstrate consistently reliable factual generation with only rare hallucinations under representative use.
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
On the ordinary reading that Meta had not achieved reliable factual generation, the claim is supported. Its strongest Llama 2 chat model scored 64.14% on the combined truthful-and-informative measure, and the technical report acknowledged hallucination limitations. This does not establish that Meta lagged competitors: its report and Anyscale’s experiment contain favorable comparisons. The grade depends on the absolute reliability reading of poor.
Limitations
- TruthfulQA measures a challenging question set, not typical user traffic.
- The combined score does not imply a 35.86% hallucination rate.
- Poor is qualitative, and a comparative interpretation would not receive this verdict.
- Anyscale sold services running the models it evaluated.
Evidence (3 items)
Supports the claim: Meta,
Llama-2-Chat 70B 64.14 0.01
Why it matters here: The 70B chat model achieved 64.14% generations judged both truthful and informative on TruthfulQA, showing a substantial gap from reliable factual answering. This result is also present in the dated July technical report; unsuccessful answers cannot all be counted as hallucinations.
- Publisher
- Meta
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Meta researchers via arXiv,
toxicity and truthfulness.
Why it matters here: The report ranks Llama 2-Chat favorably against these peers while separately acknowledging a propensity to hallucinate in section 5.2. It supports the reliability concern but weighs against reading Marcus as claiming inferior performance relative to competitors.
- Where
- Section 4.2, Truthfulness, Toxicity, and Bias; section 5.2, Limitations and Ethical Considerations
- Publisher
- Meta researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Anyscale,
are almost on par.
Why it matters here: Anyscale found strong performance by Llama-2-70b when choosing between correct and incorrect summaries. Recognizing a supplied factual summary differs from generating one without hallucinations, so this is material context rather than direct contradiction.
- Publisher
- Anyscale
- Kind of source
- Independent measurement
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, 𝐩(𝐝𝐨𝐨𝐦)
When I think about how poor his company (Meta) is even at making LLMs that don’t hallucinate, I get worried.
- Where
- paragraph 23
- Published
- 2023-08-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-d28-267622-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.