“Instead, everything that LLMs say -- true or false -- comes from the same process of statistically reconstructing what words are likely in some context. They NEVER fact-check what they say.”
Our reading, the sentence that was checked:
All LLM output, true or false, comes from the same statistical word-reconstruction process, and LLMs never fact-check what they say.
Assessment: Mixed Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Humans versus Machines: The Hallucination Edition, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-humans-versus-machines-the-hallucination-55d75c-a2
We recorded 6 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2024-04-21. 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)
From the speaker
- everything that LLMs say -- true or false
- comes from the same process of statistically reconstructing what words are likely in some context
- They NEVER fact-check what they say
Our interpretation
- Plain reading: LLM outputs are produced by next-token prediction, and LLMs never check their claims.
- He names no period. The claim describes the state of LLMs when the post appeared. The target period runs from GPT-4's release (March 2023) to the post date (2024-04-21).
- 'NEVER' is a universal claim. One documented case of an LLM, or an LLM product, checking its own claims counts against it.
Counts as supported if: Documentation shows LLMs generate text by next-token prediction and no LLM checked its own claims during the period.
Counts as contradicted if: A documented LLM method or deployed LLM feature from the period checks the model's own claims.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Mixed
Part of the claim holds and part does not, and the record says which part is which.
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 first part holds for GPT-4. Its technical report, published 15 March 2023, says it is a Transformer-based model pre-trained to predict the next token in a document, with post-training that improves factuality (e1). The 'NEVER fact-check' part does not hold under the contract's universal reading. On 19 September 2023, Google said Bard's 'Google it' button would read Bard's own response and check whether web content supports it (e2). On 20 September 2023, Meta researchers described Chain-of-Verification, in which the model plans verification questions to fact-check its own draft and then revises it (e3). Both dates fall within the target period, before the post on 21 April 2024. This result depends on counting prompted or tool-assisted checking as an LLM fact-checking what it says.
Limitations
- The packet documents next-token generation only for GPT-4 and does not cover every LLM in use during the period.
- Bard's check relies on Google Search, a tool outside the model. Under a reading limited to the bare model, it may not count.
- Chain-of-Verification needs a specific prompting method. The packet does not show it as default behavior in any deployed product.
- Prompted self-checking is itself produced by statistical generation, so the evidence does not contradict the first part of the claim.
- The packet does not show how often these checking features were used or how accurate they were.
Evidence (3 items)
Supports the claim: OpenAI (arXiv),
GPT-4 is a Transformer-based model pre-trained to predict the next token in a document.
Why it matters here: This supports the first part. The leading LLM generates text by predicting likely next tokens. The report also notes post-training, which shapes but does not replace that process.
- Where
- Abstract
- Publisher
- OpenAI (arXiv)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: Google,
When you click on the “G” icon, Bard will read the response and evaluate whether there is content across the web to substantiate it.
Why it matters here: From September 2023, Google's deployed LLM product, at the user's request, checked its own answers against web content. This contradicts 'NEVER fact-check' as applied to LLM products.
- Publisher
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Goes against the claim: arXiv (Dhuliawala et al., Meta AI),
the model first (i) drafts an initial response; then (ii) plans verification questions to fact-check its draft;
Why it matters here: Before the post, LLMs were shown fact-checking their own drafts when prompted. This contradicts the universal 'NEVER'.
- Where
- Abstract
- Publisher
- arXiv (Dhuliawala et al., Meta AI)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Humans versus Machines: The Hallucination Edition
Instead, everything that LLMs say -- true or false -- comes from the same process of statistically reconstructing what words are likely in some context. They NEVER fact-check what they say.
- Where
- paragraph 10
- Published
- 2024-04-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-humans-versus-machines-the-hallucination-55d75c-a2: Current. Outcome: Mixed. 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-humans-versus-machines-the-hallucination-55d75c-a2.