“Other models (like o3) that could actually read web links (at least to some degree) might do better, but the tendency to confabulation continues to be widespread.”
Our reading, the sentence that was checked:
Even models like o3 that can read web links may do better, but confabulation remains widespread across LLMs.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous., paragraph 39.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-llms-dishonest-unpredictable-and-d138ff-a1
We recorded 17 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 2025-06-22. 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
- Other models (like o3)
- could actually read web links (at least to some degree)
- might do better
- the tendency to confabulation continues to be widespread
Our interpretation
- Confabulation means presenting invented, unsupported, or false factual material as an answer.
- Widespread means a recurring problem across multiple leading model families and evaluation settings. It does not mean that most responses are false.
- The statement that web-capable models might do better expresses a possibility, not a promised improvement or conditional forecast.
- For this ongoing observation, use the current calendar quarter through publication, April 1–June 22, 2025, as the latest reporting window for contemporary model evaluations. Marcus supplied no reporting period.
Counts as supported if: Contemporary primary evaluations demonstrate recurring factual fabrication across models from multiple developers, including evidence that o3 remains susceptible despite possessing browsing capabilities.
Counts as contradicted if: Representative contemporary evaluations show that confabulation is confined to isolated models or exceptional cases, with consistently grounded factual output across leading model families and varied tasks.
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
Contemporary evidence supports confabulation as a recurring problem across leading LLMs. OpenAI documents substantial hallucination rates for o3 on two factual benchmarks, and Phare reports fabricated information across models from eight laboratories. Performance varies by model, task, and prompting. These findings support widespread susceptibility, not a claim that most ordinary responses are fabricated. The suggestion that browsing might help is hedged and is not scored as a forecast.
Limitations
- OpenAI's benchmark results are developer-reported and designed to elicit hallucinations.
- Phare's cited report does not provide a single population-wide hallucination rate.
- The evidence does not establish the same error rates when browsing is enabled.
Evidence (2 items)
Supports the claim: OpenAI,
more inaccurate/hallucinated claims.
Why it matters here: OpenAI reports o3 hallucination rates of 51% on SimpleQA and 33% on PersonQA, alongside nonzero rates for o4-mini and o1. The introduction establishes browsing capability, but these results do not establish performance with browsing enabled.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: Giskard Phare researchers, published on Hugging Face,
leading LLMs confidently produce factually inaccurate information.
Why it matters here: The researchers report fabricated factual details across leading models from eight laboratories, using factual questions, misinformation, debunking, and tool-use tests. They also report variation with prompts and better resistance in some models, which limits any claim of uniform failure.
- Publisher
- Giskard Phare researchers, published on Hugging Face
- 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, LLMs: Dishonest, unpredictable and potentially dangerous.
Other models (like o3) that could actually read web links (at least to some degree) might do better, but the tendency to confabulation continues to be widespread.
- Where
- paragraph 39
- Published
- 2025-06-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-llms-dishonest-unpredictable-and-d138ff-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-llms-dishonest-unpredictable-and-d138ff-a1.