“Maarten Sap from Yejin Choi’s lab dove deeper, with a more thorough investigation, and confirmed what we suspected: the claim that large language models have mastered theory of mind is a myth:”
Our reading, the sentence that was checked:
The claim that large language models have mastered theory of mind is a myth.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, GPT-5 and irrational exuberance, paragraph 42.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-gpt-5-and-irrational-exuberance-310b61-a2
We recorded 17 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 2023-04-08. 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
- Maarten Sap
- a more thorough investigation
- large language models have mastered theory of mind
- is a myth
Our interpretation
- Mastery means robust competence in reasoning about other people's beliefs and mental states across controlled tasks, rather than success on selected false-belief questions.
- This concerns contemporary LLMs, including GPT-4, rather than a permanent impossibility for future models.
- The claim also attributes relevant confirming findings to Sap and colleagues.
- I use February 2023 through the post date as the recent reporting window for the theory-of-mind debate described in the surrounding passage.
Counts as supported if: Contemporary controlled investigations, including Sap and colleagues, reveal substantial failures in mental-state reasoning that are incompatible with robust mastery.
Counts as contradicted if: Contemporary LLMs demonstrate robust generalization across varied, controlled theory-of-mind tasks, and the attributed investigation does not substantiate meaningful failures of that competence.
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
Sap and colleagues' April 3, 2023 preprint update reported 59% accuracy for GPT-4 and 60% for GPT-3.5-Turbo on ToMi mental-state questions [e5]. These substantial failures support the claim that contemporary LLMs had not mastered theory of mind and substantiate the attribution to Sap's investigation [e5]. Kosinski's March 14, 2023 preprint reported 95% for GPT-4 on 40 classic false-belief tasks [e6]. That positive result shows success on those tasks, but does not establish robust mastery across the tasks covered by both studies [e5, e6].
Limitations
- The studies test particular models, tasks and prompts, rather than every aspect of theory of mind [e5, e6].
- Sap and colleagues flag possible benchmark contamination and limited training-data transparency [e5].
- These contemporary results do not establish a permanent limitation of future LLMs [e5, e6].
Evidence (2 items)
Supports the claim: Maarten Sap, Ronan LeBras, Daniel Fried and Yejin Choi / arXiv,
surpassing GPT-4’s 59% accuracy.
Why it matters here: The extension tests contemporary models, including gpt-4-0314, and finds substantial errors on mental-state questions. These results substantiate the attribution to Sap and are incompatible with robust mastery under the stated reading.
- Publisher
- Maarten Sap, Ronan LeBras, Daniel Fried and Yejin Choi / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Michal Kosinski / arXiv,
solved nearly all the tasks (95%).
Why it matters here: Strong performance on classic false-belief tasks shows genuine positive test results that must be acknowledged. This narrower task set does not establish robustness across the mental-state tests where Sap and colleagues found failures.
- Where
- Abstract, version 3
- Publisher
- Michal Kosinski / arXiv
- 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, GPT-5 and irrational exuberance
Maarten Sap from Yejin Choi’s lab dove deeper, with a more thorough investigation, and confirmed what we suspected: the claim that large language models have mastered theory of mind is a myth:
- Where
- paragraph 42
- Published
- 2023-04-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-gpt-5-and-irrational-exuberance-310b61-a2: 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-gpt-5-and-irrational-exuberance-310b61-a2.