“This is because, despite there being a model that checks its work, the work-checking part of the model is still capable of hallucinations.”
Our reading, the sentence that was checked:
o1 is more prone to hallucinations because its work-checking component can itself hallucinate.
Assessment: Unresolved Evidence as of
Where and when Ed Zitron wrote it
Where's Your Ed At, The Subprime AI Crisis, paragraph 39.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment wyea-subprimeai-79c5cb-a1
We recorded 50 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 2024-09-16. 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
- This is because
- a model that checks its work
- the work-checking part of the model
- still capable of hallucinations
Our interpretation
- The preceding sentence supplies the comparison with GPT-4o. Assess whether errors in o1's checking process explain the asserted greater tendency to hallucinate.
- The relevant model is the September 2024 o1-preview release, not later o1 versions.
- Treat 'work-checking part' as the checking process described by the speaker; do not assume a separately deployed verifier without architecture evidence.
- Showing that reasoning can contain errors is weaker than establishing that those errors cause a comparative increase in final-answer hallucinations.
Counts as supported if: Architecture evidence identifies the relevant checking process, and controlled component experiments show that its hallucinations cause higher final-answer hallucination rates than GPT-4o on representative matched tasks.
Counts as contradicted if: Architecture or controlled experiments rule out the proposed mechanism, or robust representative comparisons refute the asserted increase. A limited benchmark can challenge the comparison without settling the causal mechanism.
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 preview card reports fewer hallucinations on its tests while acknowledging anecdotal feedback suggesting more. Zitron therefore had a basis for mentioning those reports, but they do not establish his causal explanation. The launch documentation describes learned self-correction without identifying or experimentally isolating the proposed faulty checking component. The benchmark evidence challenges the comparative premise, but is insufficient to settle the broader causal claim.
Limitations
- The evaluations were reported by the model developer and cover limited tasks.
- The archived card is dated September 12; September 17 is used conservatively as its availability date from the file identifier.
- No inference is made from later o1 model versions.
Evidence (2 items)
Goes against the claim: OpenAI,
According to these evaluations, o1-preview hallucinates less frequently than GPT-4o, and o1-mini hallucinates less frequently than GPT-4o-mini.
Why it matters here: The preview card's tests challenge the comparative premise, but the same paragraph acknowledges anecdotal feedback in the opposite direction. Neither establishes the proposed cause; the later file date is used conservatively for availability.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI,
It learns to recognize and correct its mistakes.
Why it matters here: The launch description identifies learned self-correction through reinforcement learning. It does not establish a separately identifiable checking component whose hallucinations cause a comparative increase.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Documentation identifying the checking mechanism and controlled experiments showing whether its errors cause the claimed comparative increase on matched, representative tasks.
Where it was said (1)
Said once in the sources this register covers.
: Where's Your Ed At, The Subprime AI Crisis
This is because, despite there being a model that checks its work, the work-checking part of the model is still capable of hallucinations.
- Where
- paragraph 39
- Published
- 2024-09-16
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment wyea-subprimeai-79c5cb-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.