“To a next-token predictor, a phrase like Harry Shearer lives in a particular part of a multidimensional space. Words in that space are often followed by words like “British actor”. So out comes a hallucination.”
Our reading, the sentence that was checked:
The AI's hallucination that Harry Shearer is a 'British actor' came about because, to a next-token predictor, a phrase like 'Harry Shearer' sits in a part of a multidimensional space where words are often followed by words like 'British actor'.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why DO large language models hallucinate?, paragraph 18.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-why-do-large-language-models-hallucinate-e6e804-a1
We recorded 16 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 2025-05-05. 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)
Conditions
- Harry Shearer 'British actor' example (scope, stated by the speaker)
From the speaker
- Harry Shearer
- a particular part of a multidimensional space
- often followed by words like “British actor”
- So out comes a hallucination.
Our interpretation
- The claim explains the particular erroneous biography described in the post.
- The ordinary reading is that learned associations between Shearer, entertainers and British nationality materially caused this completion. It does not require a particular mathematical definition of a neighbourhood.
Counts as supported if: An analysis of the implicated system and completion shows that the proposed associations materially contributed to the nationality error, supported by controlled interventions or comparably strong causal evidence.
Counts as contradicted if: An analysis of that completion identifies a different mechanism and shows that the proposed associations did not materially contribute.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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 explanation is testable, but the search did not find evidence tracing this particular error to the proposed associations. Research on another model establishes that hallucination mechanisms can be investigated experimentally; it does not settle this example. Neither the existence of the erroneous biography nor the next-token training objective establishes its specific cause.
Limitations
- Public searches cannot establish that no relevant internal investigation exists.
- Unresolved does not mean that Marcus's explanation is false.
Evidence (1 item)
Context only: Anthropic,
Sometimes, this sort of “misfire” of the “known answer” circuit happens naturally, without us intervening, resulting in a hallucination.
Why it matters here: Anthropic reports interventions involving entity-recognition and answer-refusal features. This demonstrates a way to investigate hallucination mechanisms, but does not establish or refute the cause of the Shearer output.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
A reproducible analysis of the implicated model and completion, including retrieval records and interventions that test the proposed nationality associations.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Why DO large language models hallucinate?
To a next-token predictor, a phrase like Harry Shearer lives in a particular part of a multidimensional space. Words in that space are often followed by words like “British actor”. So out comes a hallucination.
- Where
- paragraph 18
- Published
- 2025-05-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-why-do-large-language-models-hallucinate-e6e804-a1: Current. Outcome: Unresolved. 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-why-do-large-language-models-hallucinate-e6e804-a1.