“The problem of course is that some significant fraction of that science will be wrong, because of the inevitable tendency of LLMs to hallucinate.”
Our reading, the sentence that was checked:
Because LLMs inevitably hallucinate, a significant fraction of science produced with them will be wrong.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Slopocalypse Now, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-slopocalypse-now-60c57f-a1
We recorded 6 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-08-04. 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
- science produced with AI assistance (qualifier, stated by the speaker)
From the speaker
- some significant fraction
- that science will be wrong
- because of
- inevitable tendency of LLMs to hallucinate
Our interpretation
- Assess the causal generalization about science produced with LLM assistance; no deadline is supplied.
- AI assistance defines the population being discussed, rather than a separate conditional forecast antecedent.
- The plain reading requires a nontrivial share of scientific work to contain material falsehoods caused by LLM hallucinations.
- Treat 'inevitable tendency' as a recurring failure risk, rather than a claim that every output must be false.
- Errors in raw model answers and fabricated references are relevant, but do not alone establish the fraction of finished scientific work that is materially wrong.
Counts as supported if: Representative audits of LLM-assisted scientific work establish a substantial paper-level rate of material falsehoods and connect those errors to LLM hallucinations.
Counts as contradicted if: Representative audits with adequate detection show that hallucination-caused material falsehoods are negligible in finished LLM-assisted scientific work, including after normal verification.
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
Real scientific literature contains large numbers of apparently hallucinated references. That supports Marcus's concern, but does not establish the fraction of AI-assisted science that is materially wrong or prove the cause of each error. A scientific-agent evaluation also shows that carefully designed workflows can improve factual accuracy. The claimed population-level causal consequence remains unresolved.
Limitations
- The largest real-world audit cited here is a preprint.
- Successful benchmark performance does not establish reliability across science.
- No numerical meaning for 'significant fraction' was supplied.
Evidence (2 items)
Context only: arXiv / Zhenyue Zhao and colleagues,
146,932 hallucinated
Why it matters here: The audit supports widespread contamination of scientific references and an association with LLM adoption. It does not supply the share of AI-assisted scientific work whose substantive findings are wrong, and its attribution is statistical.
- Publisher
- arXiv / Zhenyue Zhao and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv / Michael D. Skarlinski and colleagues,
more accurate
Why it matters here: This evaluation shows that a designed LLM workflow can produce comparatively accurate scientific summaries. It limits an inference from hallucination risk to unavoidable failure in finished scientific work, without establishing a population-wide error rate.
- Publisher
- arXiv / Michael D. Skarlinski and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A representative audit of identified LLM-assisted scientific work that measures material paper-level errors, verifies their origin and accounts for human checking.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Slopocalypse Now
The problem of course is that some significant fraction of that science will be wrong, because of the inevitable tendency of LLMs to hallucinate.
- Where
- paragraph 20
- Published
- 2025-08-04
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-slopocalypse-now-60c57f-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.