GPT-4's reasoning about the physical, psychological and mathematical world will still be unreliable.

Judge GPT-4 as released (March 2023) and evaluated during its first year or so. The claim has three domains: GPT-4's reasoning about (1) the physical world, (2) the psychological world (theory of mind, intentions) and (3) the mathematical world will be unreliable, meaning it makes non-trivial errors on such tasks rather than consistently getting them right. TRUE: published evaluations or well-documented tests show significant error rates or failures in all three domains. MOSTLY TRUE: all three are shown unreliable, but in one domain the errors are fairly limited. MIXED: unreliability is documented in roughly one or two of the three domains, while GPT-4 is shown to reason reliably in the others. MOSTLY FALSE: GPT-4 is shown reliable in most of these domains, with only isolated failures. FALSE: evaluations show GPT-4 reasoning reliably in all three domains.

Set before the research began.

1. E1 Testing theory of mind in large language models and humans Nature Human Behaviour · 20 May 2024

   “Across the battery of theory of mind tests, we found that GPT-4 models performed at, or even sometimes above, human levels at identifying indirect requests, false beliefs and misdirection, but”

2. E2 Clever Hans or Neural Theory of Mind?Stress Testing Social Reasoning in Large Language Models arXiv (ar5iv rendering) · 19 February 2023

   “Nevertheless, there has been a gradual decline in the performance of subsequent models on other categories, such as other person (from 93.8% by davinci-002 to 68.8% by GPT-4),”

3. E3 Exposing the Achilles' Heel: Evaluating LLMs Ability to Handle Mistakes in Mathematical Reasoning arXiv · 31 January 2026

   “Our analysis demonstrates that most state-of-the-art LLMs, excluding GPT-4o, struggle with mistake detection and correction.”

4. E4 Clever Hans or Neural Theory of Mind?Stress Testing Social Reasoning in Large Language Models arXiv (ar5iv rendering) · 19 February 2023

   “We find that contemporary LLMs demonstrate certain N-ToM abilities, but these abilities are not robust (§4).”

5. E5 Testing theory of mind in large language models and humans Nature Human Behaviour (Strachan et al.) · 20 May 2024

   “Across the battery of theory of mind tests, we found that GPT-4 models performed at, or even sometimes above, human levels at identifying indirect requests, false beliefs and misdirection, but”

6. E6 Testing theory of mind in large language models and humans Nature Human Behaviour (Strachan et al.) · 20 May 2024

   “Irony. GPT-4 performed significantly better than human levels”

7. E7 Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models arXiv (ar5iv rendering), authors not Gary Marcus

   “and transparent access (from 66.7% by davinci-002 to 0% by GPT-4)”

8. E8 Exposing the Achilles' Heel: Evaluating LLMs Ability to Handle Mistakes in Mathematical Reasoning arXiv

   “Yet, GPT-4 performs poorly in mistake detection (0.72 vs. 0.80).”

9. E9 GPT-4 Can’t Reason (Position Paper) arXiv (ar5iv rendering); author Konstantine Arkoudas · 26 June 2023

   “Based on the results of that analysis, this paper argues that, despite the occasional flashes of analytical brilliance, GPT-4 at present is utterly incapable of reasoning.”

10. E10 GPT-4 Can’t Reason (Position Paper) arXiv (ar5iv rendering); author Konstantine Arkoudas · 26 June 2023

    “The answer to the question posed to GPT-4 below is yes, we can conclude that p5 cannot be seated either in the middle seat or on either end. GPT-4 not only fails to derive either of these conclusions, it actually arrives at the opposite conclusion, that p5 must be in the middle.”

11. E11 Clever Hans or Neural Theory of Mind? Stress Testing Social Reasoning in Large Language Models arXiv (academic preprint by external authors) · 19 February 2023

    “find that while LLMs exhibit certain N-ToM abilities, this behavior is far from being robust.”

The evidence documents failures in psychological reasoning (E2, E4, E7, E11) and mathematical or logical reasoning (E8, E9, E10), but none of the items covers physical reasoning, and some items show GPT-4 at or above human levels on certain theory-of-mind tasks (E1, E6). The criteria require all three domains to be shown unreliable for "mostly true", so the evidence does not settle that verdict.