“Just like every other large language model I have seen.”
Our reading, the sentence that was checked:
Like Galactica, every other large language model the author has seen prevaricates (fabricates content) a lot.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A Few Words About Bullshit, paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-few-words-about-bullshit-097fc5-a1
We recorded 3 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 2022-11-16. 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)
Conditions
- every other LLM the author has seen (scope, stated by the speaker)
- a lot (qualifier, stated by the speaker)
From the speaker
- every other large language model I have seen
- It prevaricates. A lot.
Our interpretation
- Prevaricates means generates fabricated or false factual content; it does not require intentional deception.
- A lot means a recurring, substantial problem, rather than an isolated mistake. Marcus supplies no numerical threshold or distribution of prompts.
- The scope is the models Marcus had encountered by publication, not every model subsequently released.
- The present-perfect phrase 'I have seen' describes accumulated experience through 2022-11-16. Its starting date is unspecified, so the target period has no asserted start.
Counts as supported if: Identify the models Marcus had encountered by publication and establish recurring fabrication in each through documented outputs or suitable evaluations.
Counts as contradicted if: Identify a model within that set and show, through sufficiently broad and relevant evaluations, that fabrication was exceptional rather than recurring. Occasional correct answers would not suffice.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
Contemporary research documents substantial false-answer problems in several major model families. That supports the concern behind Marcus's statement, but cannot establish its full scope: every model he had seen. Truthfulness also varied with training. The evidence neither verifies the universal claim nor identifies a qualifying counterexample.
Limitations
- False answers on deliberately challenging questions are not a direct estimate of fabrication in ordinary use.
- The claim leaves both its model set and its frequency threshold unspecified.
Evidence (2 items)
Context only: Lin, Hilton and Evans; arXiv / ACL,
truthful on 58% of questions
Why it matters here: The study evaluated GPT-3, GPT-Neo/J, GPT-2 and a T5-based model on questions designed to elicit common misconceptions. It establishes substantial false-answer rates on that benchmark, but neither identifies Marcus's complete model set nor measures ordinary-use fabrication frequency.
- Publisher
- Lin, Hilton and Evans; arXiv / ACL
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Ouyang and colleagues; arXiv,
improvements in truthfulness
Why it matters here: The researchers report improved truthfulness after training with human feedback. This qualifies comparisons across models and versions, but does not establish that any model Marcus saw rarely fabricated content.
- Publisher
- Ouyang and colleagues; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A reliable inventory of the models Marcus had encountered, paired with relevant evaluations showing recurring fabrication in each or a clear counterexample.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, A Few Words About Bullshit
Just like every other large language model I have seen.
- Where
- paragraph 15
- Published
- 2022-11-16
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-few-words-about-bullshit-097fc5-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-a-few-words-about-bullshit-097fc5-a1.