CoChat

The DOI matches its Crossref record; 78 of the 78 references it cites were found in a registry or an index; 24 of the 50 sources its claims cite were checked for support.

Evaluating large language models in theory of mind tasks

Open access
Kosinski, Michal · 2024 · Proceedings of the National Academy of Sciences

Abstract

Eleven large language models (LLMs) were assessed using 40 bespoke false-belief tasks, considered a gold standard in testing theory of mind (ToM) in humans. Each task included a false-belief scenario, three closely matched true-belief control scenarios, and the reversed versions of all four. An LLM had to solve all eight scenarios to solve a single task. Older models solved no tasks; Generative Pre-trained Transformer (GPT)-3-davinci-003 (from November 2022) and ChatGPT-3.5-turbo (from March 2023) solved 20% of the tasks; ChatGPT-4 (from June 2023) solved 75% of the tasks, matching the performance of 6-y-old children observed in past studies. We explore the potential interpretation of these results, including the intriguing possibility that ToM-like ability, previously considered unique to humans, may have emerged as an unintended by-product of LLMs’ improving language skills. Regardless of how we interpret these outcomes, they signify the advent of more powerful and socially skilled AI—with profound positive and negative implications.

Overview

CoChat Summary

TL;DR

This paper tests whether large language models (LLMs) can solve classic false-belief theory-of-mind (ToM) tasks, using 40 bespoke tasks of two types and multiple controls. The main result is that performance improved with model scale, with ChatGPT-4 solving 75% of tasks, though the authors stop short of claiming this settles whether LLMs “really” have ToM.

Key findings

  • Across 11 LLMs, performance improved over time: older models up to 2022 failed false-belief scenarios or controls in all tasks, GPT-3-davinci-002 solved 5%, GPT-3-davinci-003 and ChatGPT-3.5-turbo solved 20%, and ChatGPT-4 solved 75% of tasks.
  • ChatGPT-4 performed especially well on the two task families: it solved 90% of Unexpected Contents tasks and 60% of Unexpected Transfer tasks.
  • Performance on the two task types was strongly correlated, with reported correlation R = 0.98 and 95% CI [0.92, 0.99].
  • In the sentence-by-sentence analyses, ChatGPT-4’s probability of the correct belief prediction updated as story information accumulated; for example, in the Unexpected Contents task, after the false label was revealed and Sam did not look inside, the model became certain Sam’s belief was false.

Methods

  • The authors created 20 Unexpected Contents tasks and 20 Unexpected Transfer tasks, each with a false-belief scenario, three true-belief controls, and reversed versions of all four scenarios; solving one task required correctly answering 16 prompts across eight scenarios.
  • They tested 11 LLMs: GPT-1, GPT-2, six GPT-3-family models, ChatGPT-3.5-turbo, ChatGPT-4, and Bloom.
  • Each story was followed by prompts assessing both the actual state of affairs and the protagonist’s belief; responses were evaluated over many completions using API calls, with temperature = 0 for replicability in some demonstrations and 1,000 completions for probability estimates in others.

Limitations

  • The paper states that results should be taken “with much skepticism,” and that failures do not necessarily prove lack of ToM, while successes do not automatically prove ToM either.
  • It notes several evident confounds: models may rely on guessing, memorization, response patterns, or task-specific shortcuts rather than mental-state tracking.
  • The paper also says that some answers were hard to score because models sometimes gave partially correct or context-dependent responses that did not fit the scoring key, which could underestimate performance.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1073/pnas.2405460121

Cites

78 of 78 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. Changes in the probabilities of ChatGPT-4’s completions of Prompts 1.1 and 1.2 as the story was revealed in one-sentence increments.
Fig. 1. Changes in the probabilities of ChatGPT-4’s completions of Prompts 1.1 and 1.2 as the story was revealed in one-sentence increments.CC BY-NC-ND · PubMed Central
Fig. 2. Changes in the probabilities of ChatGPT-4’s completions of Prompts 2.1 and 2.2 as the story was revealed to it in one-sentence increments. The last sentence of the story (“John comes back home and wants to play with the cat.”) was added to Prompt 2.2, as this prompt made little sense on its own throughout most of the story.
Fig. 2. Changes in the probabilities of ChatGPT-4’s completions of Prompts 2.1 and 2.2 as the story was revealed to it in one-sentence increments. The last sentence of the story (“John comes back home and wants to play with the cat.”) was added to Prompt 2.2, as this prompt made little sense on its own throughout most of the story.CC BY-NC-ND · PubMed Central
Fig. 3. The percentage of false-belief tasks solved by LLMs (out of 40). Each task contained a false-belief scenario, three accompanying true-belief scenarios, and the reversed versions of all four scenarios. A model had to solve 16 prompts across all eight scenarios to score a single point. The number of parameters and models’ publication dates are in parentheses. The number of parameters for models in the GPT-3 family was estimated by Gao (55) and for ChatGPT-4 by Patel and Wong (56). Average children’s performance on false-belief tasks was reported after a meta-analysis of 178 studies (54). Error bars represent 95% CI.
Fig. 3. The percentage of false-belief tasks solved by LLMs (out of 40). Each task contained a false-belief scenario, three accompanying true-belief scenarios, and the reversed versions of all four scenarios. A model had to solve 16 prompts across all eight scenarios to score a single point. The number of parameters and models’ publication dates are in parentheses. The number of parameters for models in the GPT-3 family was estimated by Gao (55) and for ChatGPT-4 by Patel and Wong (56). Average children’s performance on false-belief tasks was reported after a meta-analysis of 178 studies (54). Error bars represent 95% CI.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

15 Valid · 5 Not valid · 8 Not checked · 100 Own results and interpretations

56 more claims

Create a free account to see all 128 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 34
  • Probably not a substantial claim. 47
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 68 references

Research works printed but not linked: 4

  1. A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, Improving language understanding by generative pre-training. OpenAI (2018). https://openai.com/index/language-unsupervised/. Accessed 1 August 2023.
  2. R. Alec , Language models are unsupervised multitask learners. OpenAI Blog 1 (2019). https://api.semanticscholar.org/CorpusID:160025533. Accessed 1 February 2023.
  3. A. Vaswani “Attention is all you need” in Proceedings of the 31st International Conference on Neural Information Processing Systems I. Guyon Eds. (Curran Associates Inc. 2017) pp. 6000–6010.
  4. N. Block, Troubles with functionalism. Minn. Stud. Philos. Sci. 9 261–325 (1978).

Cited by

See all 339

Metadata

Authors
Kosinski, M.
Affiliations
Graduate School of Business, Stanford University, Stanford, CA 94305
Venue
Proceedings of the National Academy of Sciences
Published
2024-10-29
Volume
121
Issue
45
Pages
e2405460121-e2405460121
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
OpenAlex
W4403863303
Cited by
339
References
68
Citations by year
2026: 113 · 2025: 112 · 2024: 57
FWCI
219.35
Type
journal article
Language
en
Open access
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
  • arXiv (Cornell University)
Topics
  • Social Sciences › Psychology › Developmental and Educational Psychology › Child and Animal Learning Development
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Social Sciences › Psychology › Social Psychology › Social Robot Interaction and HRI
Keywords
large language models, theory of mind, false belief, false belief tasks, GPT-4, ChatGPT, child cognitive development
MeSH terms
Child, Female, Humans, Language, Male, Theory of Mind

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Evaluating large language models in theory of mind tasks. https://app.cochat.ai/papers/10.1073/pnas.2405460121
The paper
Kosinski, M. (2024). Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2405460121

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.