“The very process that LLMs use to generalize also creates hallucinations.”
Our reading, the sentence that was checked:
The same process LLMs use to generalize is what creates their hallucinations.
Assessment: Supported 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-47c312-a1
We recorded 6 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
- The very process that LLMs use to generalize
- also creates hallucinations
- even with perfect data, the stochastic reconstructive process would still produce some errors
Our interpretation
- Plain reading: hallucinations are a byproduct of the same statistical learning and prediction that lets LLMs generalize, not a separate bug or a result of bad data.
- The claim describes LLMs in general, so no target period is set.
Counts as supported if: Research shows that statistical pretraining and prediction produce hallucinations, including with error-free training data.
Counts as contradicted if: Research shows hallucinations come mainly from a separate mechanism, such as bad data alone, and not from the generalizing prediction process.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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
A peer-reviewed STOC 2024 paper shows that pretrained models built to predict well must hallucinate certain facts at a minimum rate. This has nothing to do with architecture or data quality, and holds even with ideal training data (e1). OpenAI's 2025 write-up likewise says pattern-based next-word prediction cannot get arbitrary rare facts right, and so produces hallucinations (e2). Both match Marcus's claim that the process LLMs use to generalize also creates hallucinations, even with perfect data. One caveat: OpenAI says models can be trained to abstain when unsure, so hallucinations can be reduced (e3). That does not undercut the claim about where they come from.
Limitations
- The theory covers rare, arbitrary facts most clearly. Other kinds of hallucination may have other causes.
- The supporting authors overlap (Kalai and Vempala are on both papers).
Evidence (3 items)
Supports the claim: arXiv / STOC 2024 (Kalai and Vempala),
there is an inherent statistical lower-bound on the rate that pretrained language models hallucinate certain types of facts, having nothing to do with the transformer LM architecture or data quality.
Why it matters here: Hallucination follows from statistical prediction itself, even with ideal data. That matches the claim that the generalizing process creates hallucinations.
- Where
- Abstract
- Publisher
- arXiv / STOC 2024 (Kalai and Vempala)
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Supports the claim: OpenAI,
Spelling and parentheses follow consistent patterns, so errors there disappear with scale. But arbitrary low-frequency facts, like a pet’s birthday, cannot be predicted from patterns alone and hence lead to hallucinations.
Why it matters here: OpenAI traces hallucinations to the pattern-based next-word prediction that makes models fluent, which supports the claim.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI,
Claim: Hallucinations are inevitable. Finding: They are not, because language models can abstain when uncertain.
Why it matters here: This adds a caveat: post-training to abstain can reduce hallucinations. It does not deny that pretraining prediction creates them.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Humans versus Machines: The Hallucination Edition
The very process that LLMs use to generalize also creates hallucinations.
- 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-47c312-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-humans-versus-machines-the-hallucination-47c312-a1.