“The chronic problem with creating fake citations in research papers and faked cases in legal briefs is a manifestation of the same problem; LLMs correctly “model” the structure of academic references, but often make up titles, page numbers, journals and so on — once again failing to sanity check their outputs against information (in this case lists of references) that are readily found on the internet.”
Our reading, the sentence that was checked:
Fake citations and faked legal cases from LLMs stem from the same problem of mimicking structure without sanity-checking against available information.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why DO large language models hallucinate?, paragraph 21.
- 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-df0a1b-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)
From the speaker
- fake citations in research papers and faked cases in legal briefs
- a manifestation of the same problem
- correctly “model” the structure of academic references
- often make up titles, page numbers, journals
- failing to sanity check their outputs against information
Our interpretation
- Read this as a causal explanation: generating plausible reference structures without dependable verification materially produces fabricated scholarly and legal citations.
- The claim does not require identical neural circuits in every model or assert that verification failure is the only contributing cause.
- Separate the documented occurrence of fabricated references from evidence establishing the proposed shared cause.
Counts as supported if: Studies establish plausible but fabricated references and show that generation without source verification materially causes such errors in both scholarly and legal settings, consistent with the proposed structural-imitation mechanism.
Counts as contradicted if: Relevant controlled studies show that the proposed mechanism does not materially explain the fabrications, or establish incompatible causes for the claimed shared explanation.
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 descriptive core is well documented: models generated recognisable reference formats containing invented works, and ChatGPT-generated fake cases reached court. OpenScholar also supports the practical value of grounding and feedback. These findings are consistent with Marcus's explanation, but they do not establish that structural imitation without verification is the shared cause across both settings. The causal proposition therefore remains unresolved; the fabrication problem itself does not.
Limitations
- Successful mitigation is evidence about contributors, but a combined system change does not isolate every causal component.
- The evidence does not justify a universal claim that all LLM systems lack verification.
Evidence (3 items)
Context only: Scientific Reports,
With both GPT-3.5 and GPT-4, every citation is in APA format.
Why it matters here: The study examined 636 citations generated in April 2023: 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated despite recognisable formatting. This supports the described symptom, but does not independently identify the causal mechanism.
- Publisher
- Scientific Reports
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: U.S. District Court, Southern District of New York; CourtListener copy,
they submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT
Why it matters here: The order documents fabricated authorities in the March 1 filing and subsequent failures to verify them. It establishes the legal incident, but contains no investigation of the model's internal cause.
- Publisher
- U.S. District Court, Southern District of New York; CourtListener copy
- Kind of source
- Regulatory filing
- Describes events of
- Public from
Context only: Asai and colleagues, arXiv,
While GPT4o hallucinates citations 78 to 90% of the time, OpenScholar achieves citation accuracy on par with human experts.
Why it matters here: The retrieval and feedback system supplies evidence that grounding can improve scholarly citation accuracy. The combined system comparison does not isolate structural imitation as the cause or establish the same mechanism for legal fabrications.
- Where
- Abstract
- Publisher
- Asai and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled scholarly and legal citation experiments that isolate source verification and analyse how plausible reference structures become fabricated content.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Why DO large language models hallucinate?
The chronic problem with creating fake citations in research papers and faked cases in legal briefs is a manifestation of the same problem; LLMs correctly “model” the structure of academic references, but often make up titles, page numbers, journals and so on — once again failing to sanity check their outputs against information (in this case lists of references) that are readily found on the internet.
- Where
- paragraph 21
- Published
- 2025-05-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-why-do-large-language-models-hallucinate-df0a1b-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-df0a1b-a1.