“Instead, LLM "hallucinations" arise, regularly, because (a) they literally don't know the difference between truth and falsehood, (b) they don't have reliably reasoning processes to guarantee that their inferences are correct and (c) they are incapable of fact-checking their own work.”
Our reading, the sentence that was checked:
LLM hallucinations arise regularly because LLMs literally don't know the difference between truth and falsehood, don't have reasoning processes reliable enough to guarantee their inferences are correct, and are incapable of fact-checking their own work.
Assessment: Mixed Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Humans versus Machines: The Hallucination Edition, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-humans-versus-machines-the-hallucination-d1a01e-a1
We recorded 7 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 2024-04-21. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- LLM "hallucinations" arise, regularly, because
- literally don't know the difference between truth and falsehood
- don't have reliably reasoning processes to guarantee that their inferences are correct
- incapable of fact-checking their own work
Our interpretation
- Assess models and methods available by the post date, 2024-04-21. This is a claim about existing capabilities and causes, not a forecast about future models.
- Read knowing the difference behaviorally: distinguishing true from false claims or accurately assessing whether an answer is correct. This does not require consciousness.
- Read fact-checking their own work as checking factual assertions in an initial response and identifying or correcting errors. Prompted verification using the same model qualifies; it need not be automatic or infallible.
- Use tests without external factual feedback for the strongest comparison with the claimed intrinsic incapacity.
- Assess the components separately. Evidence of unreliable reasoning supports that limitation, but does not by itself establish the entire causal explanation.
Counts as supported if: Controlled evaluations show unreliable reasoning and factual self-evaluation, while interventions link those failures to hallucinations. Full support for the categorical components would require evidence of inability, rather than merely imperfect performance.
Counts as contradicted if: Reproducible truth discrimination or successful verification of a model's own factual responses contradicts the categorical absence of those abilities. Successful self-verification that reduces hallucinations also challenges the proposed explanation based on incapacity.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Mixed
Part of the claim holds and part does not, and the record says which part is which.
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 reasoning-reliability limitation holds in the cited tests: intrinsic correction could reduce accuracy. But the categorical claims go further than the evidence supports. Before the post, experiments demonstrated truth assessment and a method that used the same model to check its own factual responses and improve them without external tools. Those findings contradict incapacity under the stated behavioral reading. They do not establish infallibility, human-like knowledge, or a complete alternative explanation of hallucinations. The three-part causal explanation remains unproven.
Limitations
- The verdict separates supported reliability limitations from contradicted categorical incapacity; it does not validate the full causal account.
- Results depend on models, tasks, prompts, and the meaning of fact-checking.
- All decisive studies concern capabilities demonstrated before the post.
Evidence (3 items)
Supports the claim: arXiv; Huang et al.,
the accuracies of all models drop
Why it matters here: In the reported intrinsic-correction experiments, models had to judge their answers without ground-truth feedback, and accuracy fell. This supports unreliable reasoning and self-correction in those settings, not universal incapacity or the full explanation of factual hallucinations.
- Publisher
- arXiv; Huang et al.
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: arXiv; Dhuliawala et al.,
CoVe brings improvements
Why it matters here: The model generated verification questions, answered them using the model itself without external tools, and revised its response. The reported improvements contradict categorical inability under the stated behavioral reading, while leaving substantial errors.
- Publisher
- arXiv; Dhuliawala et al.
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: arXiv; Kadavath et al.,
well-calibrated
Why it matters here: The study reports calibrated truth assessments under suitable formatting and evaluates models' assessments of their own proposed answers. This challenges the absolute absence of truth discrimination; it does not establish human-like knowledge or reliable checking in every setting.
- Publisher
- arXiv; Kadavath et al.
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled interventions isolating truth discrimination, reasoning reliability, and self-verification would clarify their separate causal contributions. A definition of knowledge and fact-checking would resolve the remaining interpretive ambiguity.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Humans versus Machines: The Hallucination Edition
Instead, LLM "hallucinations" arise, regularly, because (a) they literally don't know the difference between truth and falsehood, (b) they don't have reliably reasoning processes to guarantee that their inferences are correct and (c) they are incapable of fact-checking their own work.
- Where
- paragraph 10
- Published
- 2024-04-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-humans-versus-machines-the-hallucination-d1a01e-a1: Current. Outcome: Mixed. 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-humans-versus-machines-the-hallucination-d1a01e-a1.