“Humans can learn rules of games like chess from modest amounts of explicit instruction; LLMs can’t stick to explicit rules.”
Our reading, the sentence that was checked:
LLMs cannot stick to explicit rules such as the rules of chess.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, No, Virginia, AGI is not imminent, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-no-virginia-agi-is-not-imminent-aa764b-a3
We recorded 20 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 2023-08-20. 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
- Humans can learn rules of games like chess
- modest amounts of explicit instruction
- LLMs can’t stick to explicit rules
Our interpretation
- The plain reading is a criticism of dependable rule adherence over an interaction, not a claim that LLMs never produce legal moves.
- Chess provides the concrete test: making legal moves throughout games when asked to play, without an external engine enforcing legality.
- For this ongoing observation, I use August 2023 through publication as the latest reporting window. The monthly boundary is an evaluator convention because model evaluations have no fixed reporting cycle.
- The normalized proposition concerns LLM rule adherence. The separate claim about how quickly humans learn is not established by chess transcripts alone.
Counts as supported if: Contemporaneous tests show recurring violations of game rules or explicit interaction instructions despite clear requests to play, including failures beyond isolated notation mistakes.
Counts as contradicted if: Contemporaneous, repeated tests show dependable adherence throughout varied games and explicitly specified rule tasks under modest instruction.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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 August 19, 2023 report documents two conversationally prompted GPT-3.5 games in which it could no longer produce a legal move by moves 8 and 14, respectively [e1]. The same report records successful GPT-4 games ending in checkmate in 22 and 25 moves [e1]. These demonstrations support the claim as a criticism of dependable rule adherence, while showing that complete legal games were possible [e1].
Limitations
- The demonstrations do not establish a representative failure rate across models, prompts or games [e1, e2].
- They do not test comparative human and LLM learning of unfamiliar rules from modest instruction [e1, e2].
- Successful games preclude interpreting the finding as categorical inability to follow chess rules [e1, e2].
Evidence (2 items)
Supports the claim: Adam Yedidia, LessWrong,
unable to produce a legal move
Why it matters here: The accompanying demonstrations also document illegal GPT-3.5 moves with specialized prompting and a GPT-4 failure to obey the instruction to output only its own move. The same report includes successful legal games, so it supports unreliable adherence rather than total inability.
- Publisher
- Adam Yedidia, LessWrong
- Kind of source
- Demonstration
- Describes events of
- Public from
Context only: David Chudzicki,
plays legally the whole time
Why it matters here: A complete legal GPT-4 game shows that categorical inability would be too strong. The author also reports a prompt-dependent failure, but does not measure a representative completion rate.
- Publisher
- David Chudzicki
- Kind of source
- Demonstration
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, No, Virginia, AGI is not imminent
Humans can learn rules of games like chess from modest amounts of explicit instruction; LLMs can’t stick to explicit rules.
- Where
- paragraph 20
- Published
- 2023-08-20
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-no-virginia-agi-is-not-imminent-aa764b-a3: Current. Outcome: Supported. 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-no-virginia-agi-is-not-imminent-aa764b-a3.
HeldBy this site
The final-text check failed: citations_correct.
About assessment marcus-no-virginia-agi-is-not-imminent-aa764b-a2, which was not published.