“(b) the difficulty of large language models in distinguishing truth from falsehoods;”
Our reading, the sentence that was checked:
Neglecting reference, cognitive models and compositionality makes it hard for LLMs to tell truth from falsehood.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Three ideas from linguistics that everyone in AI should know, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-three-ideas-from-linguistics-that-8c1485-a2
We recorded 26 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 2022-06-22. 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
- caused by inadequate attention to reference, cognitive models, compositionality (qualifier, stated by the speaker)
From the speaker
- inadequate attention to these three factors
- the difficulty of large language models
- distinguishing truth from falsehoods
Our interpretation
- The surrounding introduction makes this a causal claim about reference, cognitive models and compositionality, not merely an observation that models sometimes give false answers.
- Distinguishing truth from falsehood means reliably selecting or identifying factually correct statements against plausible false alternatives.
- False-answer generation is relevant but does not by itself identify an inability to discriminate truth or its cause.
- The causal phrase is not an antecedent of a conditional forecast.
Counts as supported if: Controlled evidence shows that deficiencies in the specified capabilities materially impair true-versus-false discrimination in contemporary models, and that targeted interventions improve it.
Counts as contradicted if: Well-controlled tests rule out a material causal contribution from those deficiencies, or establish an opposite effect, on representative truth-discrimination tasks.
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
TruthfulQA, introduced in 2021 and revised in May 2022, reported truthful answers on 58% of questions for the best tested model, compared with 94% for humans [e8]. The March 2022 GopherCite study reported no improvement under TruthfulQA's truthfulness measure despite producing answers with supporting evidence [e9]. Neither study isolates the causal contribution of reference, cognitive models or compositionality, so the proposed explanation remains unresolved [e8, e9].
Limitations
- TruthfulQA's questions target misconceptions; its results do not establish an overall factual-error rate [e8].
- The cited results concern answer generation rather than a direct test of classifying supplied statements as true or false [e8, e9].
- GopherCite does not separately manipulate the three proposed capabilities or rule out their causal contribution [e9].
Evidence (2 items)
Context only: arXiv; Stephanie Lin, Jacob Hilton and Owain Evans,
truthful on 58% of questions
Why it matters here: TruthfulQA documents false-answer generation in GPT-family and other models on questions designed to elicit misconceptions. It does not manipulate reference, cognitive models or compositionality, so the reported failures do not establish the causal claim.
- Publisher
- arXiv; Stephanie Lin, Jacob Hilton and Owain Evans
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: DeepMind; arXiv,
no improvements in truthfulness
Why it matters here: GopherCite adds retrieved supporting evidence and preference-based training, yet does not improve this truthfulness measure. This cautions against equating source grounding with truth, but does not refute the broader claim involving all three linguistic factors.
- Publisher
- DeepMind; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled experiments on models available by June 22, 2022, that separately vary the proposed capabilities and measure discrimination between independently verified true and false statements while controlling supervision and training objectives.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Three ideas from linguistics that everyone in AI should know
(b) the difficulty of large language models in distinguishing truth from falsehoods;
- Where
- paragraph 10
- Published
- 2022-06-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-three-ideas-from-linguistics-that-8c1485-a2: 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-three-ideas-from-linguistics-that-8c1485-a2.