“The catch is that agents like OpenClaw are built on a foundation of LLMs, and as we well know, LLMs hallucinate and make all kinds of hard to predict and sometimes hard to detect errors.”
Our reading, the sentence that was checked:
Agents like OpenClaw are built on a foundation of LLMs, and LLMs hallucinate and make all kinds of hard-to-predict and sometimes hard-to-detect errors.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, OpenClaw (a.k.a. Moltbot) is everywhere all at once, and a disaster waiting to happen, paragraph 17.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-openclaw-aka-moltbot-is-everywhere-247be2-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 2026-02-01. 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)
Conditions
- sometimes (hard to detect) (qualifier, stated by the speaker)
From the speaker
- agents like OpenClaw
- built on a foundation of LLMs
- LLMs hallucinate
- hard to predict
- sometimes hard to detect
Our interpretation
- The ordinary reading is that LLM-based agents can produce plausible but incorrect outputs, and reliable identification of their errors remains difficult. It does not require every model or response to fail.
- For this ongoing observation, use January 2026, the latest complete calendar month before the post, to assess the publicly documented state of the technology.
- Difficulty predicting errors means that available methods do not reliably identify all forthcoming failures. It does not mean that error risk is wholly unpredictable.
Counts as supported if: Contemporaneous OpenClaw documentation establishes its use of language models, and evaluations document hallucinations across agent tasks together with substantial remaining difficulty identifying or anticipating errors.
Counts as contradicted if: OpenClaw does not rely on language models, or representative evaluations establish that the described hallucination and error-identification problems do not occur.
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
January documentation establishes OpenClaw's use of language models. AgentHallu documents varied hallucinations and substantial difficulty identifying their origins. HALT shows that error risk can nevertheless be estimated usefully. Together these support the ordinary claim of fallibility and sometimes difficult detection, without supporting absolute unpredictability.
Limitations
- The research is preprint evidence.
- Attribution difficulty is distinct from predicting a failure before it happens.
- The result does not establish a failure rate for OpenClaw or every supported model.
Evidence (3 items)
Supports the claim: OpenClaw project,
models OpenClaw can use
Why it matters here: The versioned documentation describes primary and fallback models and provides Claude model configurations. It establishes the language-model foundation of the agent.
- Publisher
- OpenClaw project
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: Xuannan Liu and colleagues, arXiv,
41.1% step localization accuracy
Why it matters here: AgentHallu documents hallucinations across seven agent frameworks and five domains. Its detection and attribution results support the claim that identifying errors can be difficult; step localization is not a prospective prediction score.
- Publisher
- Xuannan Liu and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Rohan Bhatnagar and colleagues, arXiv,
strong out-of-distribution generalization
Why it matters here: HALT reports useful uncertainty signals and improved hallucination screening across several model families. This limits any suggestion that errors are entirely unpredictable, while the reported discrimination remains imperfect.
- Publisher
- Rohan Bhatnagar 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, OpenClaw (a.k.a. Moltbot) is everywhere all at once, and a disaster waiting to happen
The catch is that agents like OpenClaw are built on a foundation of LLMs, and as we well know, LLMs hallucinate and make all kinds of hard to predict and sometimes hard to detect errors.
- Where
- paragraph 17
- Published
- 2026-02-01
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-openclaw-aka-moltbot-is-everywhere-247be2-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-openclaw-aka-moltbot-is-everywhere-247be2-a1.