“They do not sanity check their own work.”
Our reading, the sentence that was checked:
LLMs do not sanity check their own work.
Assessment: Contradicted Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, LLMs are not like you and me—and never will be., paragraph 7.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-llms-are-not-like-you-and-meand-never-db6a26-a2
We recorded 10 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 2025-08-12. 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
- They do not sanity check
- their own work
Our interpretation
- The ordinary reading is a categorical denial that LLMs examine their own reasoning or responses for mistakes.
- A counterexample must show actual checking or correction, rather than merely a claim by the model that it checked.
- Checking need not be infallible. An external verifier supplying the correction does not qualify by itself.
- Assess capabilities demonstrated by 2025-08-12. Earlier demonstrations remain relevant to this categorical capability claim; no lower time bound is asserted.
Counts as supported if: Representative evaluations of contemporary LLMs, including reasoning models, show no genuine inspection or correction of their own work without externally supplied error judgments.
Counts as contradicted if: A documented pre-post LLM examines and corrects its own reasoning or response without an external error judgment at inference time.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Contradicted
The dated evidence goes against 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 September 19, 2024 SCoRe study documented models correcting their own mistakes without external error judgments at inference time, including occasional correction within a single response [e1]. This provides a counterexample to the categorical denial of self-checking before August 12, 2025 [e1]. DeepSeek’s January 22, 2025 report also described spontaneous verification and reflection [e2]. Earlier experiments published October 3, 2023 found failures of intrinsic self-correction, limiting confidence in its reliability without negating the later demonstrations [e1, e3].
Limitations
- SCoRe used specialized training; its results do not establish that every model checks every answer [e1].
- The DeepSeek evidence is developer-reported and does not establish general reliability [e2].
- Demonstrated self-correction does not imply dependable error detection across tasks [e1, e3].
Evidence (3 items)
Goes against the claim: arXiv; Aviral Kumar and colleagues,
self-correct within a turn
Why it matters here: SCoRe evaluates correction without external feedback at inference time and documents revision of the model's own mistakes. It supplies a behavioral counterexample, although the models received specialized training; the event date denotes the capability documented by publication.
- Publisher
- arXiv; Aviral Kumar and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: arXiv; DeepSeek-AI,
self-verification, reflection,
Why it matters here: The report describes the model spontaneously revisiting earlier reasoning during generation. This supports the existence of self-checking, although the developer's report does not establish its general reliability; the date is when this capability was publicly documented in the cited version.
- Publisher
- arXiv; DeepSeek-AI
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Jie Huang and colleagues,
not yet capable of self-correcting their reasoning.
Why it matters here: The earlier experiments found that intrinsic revision could fail or reduce accuracy. This supports caution about dependable self-correction, but does not negate later demonstrations or prove that models never check their work.
- Publisher
- arXiv; Jie Huang and colleagues
- 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, LLMs are not like you and me—and never will be.
They do not sanity check their own work.
- Where
- paragraph 7
- Published
- 2025-08-12
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-llms-are-not-like-you-and-meand-never-db6a26-a2: Current. Outcome: Contradicted. 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-are-not-like-you-and-meand-never-db6a26-a2.