“It’s the same problem: correlations are fragile, and you can’t count on them.”
Our reading, the sentence that was checked:
GPT's failure to stick to the rules of chess, even when it plays a decent game, reflects the same problem: correlations are fragile and cannot be counted on.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Muddles about Models, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-muddles-about-models-566f8e-a1
We recorded 8 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 2023-10-05. 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)
Conditions
- in the context of GPT playing chess (scope, stated by the speaker)
From the speaker
- correlations are fragile
- you can’t count on them
- doesn’t really model the rules of chess well enough to stick to them
Our interpretation
- The claim concerns the cause of rule violations by GPT systems contemporary with the October 2023 post.
- The plain reading is that reliance on learned correlations explains recurring illegal moves. Demonstrating illegal moves alone does not establish this explanation.
- Controlled changes to board-state information, training, or internal representations could distinguish this explanation from other causes.
Counts as supported if: Contemporary GPT chess evaluations establish recurring rule violations, and controlled experiments identify reliance on correlations without adequate board-and-rule representations as a cause.
Counts as contradicted if: Controlled experiments on the relevant systems establish usable board-and-rule representations and show that the attributed absence of those representations does not explain the violations.
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
Contemporary experiments substantiate serious chess-rule failures. They do not establish Marcus’s explanation of those failures. Later intervention research also shows that next-token training can produce usable board representations in specialized models. Neither result settles the causal claim about contemporary GPT.
Limitations
- A failure to follow rules does not by itself reveal the model’s internal representations.
- The counterevidence concerns specialized later models, not the same GPT system.
Evidence (2 items)
Context only: arXiv; Mu-Tien Kuo, Chih-Chung Hsueh, and Richard Tzong-Han Tsai,
only 1.59% of games concluded naturally
Why it matters here: The experiments document substantial problems following chess rules in GPT-3.5-turbo-0301. They support the behavioral premise, but do not isolate reliance on correlations as its cause.
- Publisher
- arXiv; Mu-Tien Kuo, Chih-Chung Hsueh, and Richard Tzong-Han Tsai
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Adam Karvonen,
world model
Why it matters here: The author reports interventions in chess-model activations that change moves in line with altered board states. This challenges a blanket opposition between statistical training and internal models, but does not settle the cause of failures in the GPT systems Marcus discussed.
- Where
- Section Implications
- Publisher
- Adam Karvonen
- Kind of source
- Independent measurement
- Describes events of
- Public from
What would settle it
Controlled experiments on preserved contemporary GPT versions that measure board representations and test whether changing those representations explains or removes recurring illegal moves.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Muddles about Models
It’s the same problem: correlations are fragile, and you can’t count on them.
- Where
- paragraph 20
- Published
- 2023-10-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-muddles-about-models-566f8e-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.