“This all happens not because LLMs are smart, in the sense of being able to compare what they say with the world, but simply because they are good at making new sentences that sound old sentences, the essence of autocomplete.”
Our reading, the sentence that was checked:
LLMs produce plausible falsehoods because they mimic existing sentences like autocomplete rather than checking what they say against the world.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Seven lies in four sentences, paragraph 16.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-seven-lies-in-four-sentences-c392e1-a1
We recorded 6 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-10-04. 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)
From the speaker
- This all happens
- not because LLMs are smart
- being able to compare what they say with the world
- simply because
- making new sentences that sound old sentences
- the essence of autocomplete
Our interpretation
- Autocomplete is read as learned next-token prediction, not literal copying of stored sentences.
- The causal claim is that plausible language generation without effective factual verification explains the cited falsehoods.
- Simply makes this stronger than the observation that next-token prediction is part of model training.
- The relevant systems are those contemporary with the October 4, 2023 post.
- A base model's learned behavior and a deployed assistant's search or verification tools must be distinguished. Marcus does not specify which level his denial of comparison covers.
- Smart is assessed only through the factual-comparison capability he supplies, not as an unrestricted philosophical judgment.
Counts as supported if: Controlled evidence on relevant contemporary systems attributes plausible falsehoods to language prediction without effective factual verification, while adequately testing competing explanations.
Counts as contradicted if: Controlled evidence refutes that causal attribution, or establishes factual-comparison behavior incompatible with the claim under its stated scope.
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
Next-token pretraining supports the analogy, but it does not establish why these particular falsehoods occurred or that this was the sole explanation. Google had already documented a feature that checked generated answers against web information, which complicates the blanket contrast with factual comparison. That feature also does not establish what an unaided model can do. The evidence does not settle the causal claim without choosing a scope Marcus did not specify.
Limitations
- Training descriptions establish design facts, not a complete causal explanation of errors.
- A verification feature can exist and still fail.
- Evidence about another assistant is relevant to the general wording but cannot diagnose the specific Bing episode.
Evidence (2 items)
Context only: OpenAI; arXiv,
predict the next token
Why it matters here: This supports the factual premise behind the autocomplete analogy. The report also describes post-training alignment; the pretraining objective does not by itself establish an exclusive cause of hallucinations.
- Where
- Abstract, page 1
- Publisher
- OpenAI; arXiv
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Google,
double-check its answers.
Why it matters here: Google describes checking generated responses against supporting or contradicting web information before Marcus's post. This challenges a blanket statement about deployed assistants lacking verification, but does not prove that a bare model independently checks reality or explain the cited Bing failure.
- Publisher
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Controlled evaluations of the relevant contemporary systems that distinguish learned generation, retrieval, and verification, together with sufficient records to test the cause of the cited failures.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Seven lies in four sentences
This all happens not because LLMs are smart, in the sense of being able to compare what they say with the world, but simply because they are good at making new sentences that sound old sentences, the essence of autocomplete.
- Where
- paragraph 16
- Published
- 2023-10-04
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-seven-lies-in-four-sentences-c392e1-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-seven-lies-in-four-sentences-c392e1-a1.