“In a few cases GPT-4 set up the calculations correctly and the plug-in returned the correct answer, but then GPT-4 either misunderstood or refused to accept that answer.”
Our reading, the sentence that was checked:
In a few test cases GPT-4 set up the calculations correctly and the plug-in returned the correct answer, but GPT-4 then either misunderstood or refused to accept that answer.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Getting GPT to work with external tools is harder than you think, paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-getting-gpt-to-work-with-external-0a7866-a1
We recorded 16 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 2023-10-23. 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)
Conditions
- in a few cases (qualifier, stated by the speaker)
From the speaker
- In a few cases
- set up the calculations correctly
- the plug-in returned the correct answer
- misunderstood or refused to accept that answer
Our interpretation
- A few requires at least two distinct test cases.
- Each qualifying case must contain the complete sequence: correct setup, correct tool result, then mistaken interpretation or rejection.
- A tool error, an incorrect setup, or inability to complete a unit conversion does not by itself meet that sequence.
- The claim concerns the June–August 2023 experiment.
Counts as supported if: At least two distinct cases verify all three stages of the claimed sequence.
Counts as contradicted if: A sufficiently complete review establishes that fewer than two cases satisfy the sequence.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
The matrix-probability example clearly supports misunderstanding a correct result. Other candidates are less secure: the prime-probability tool expression was left unchecked, and the divergent-integral transcript reports tool errors rather than an unambiguous correct answer. The evidence therefore does not yet establish several qualifying cases, but it also does not disprove their existence.
Limitations
- A 2024 transcript revision supplies important qualifications about the 2023 tests.
- A correct intermediate result followed by failure is not automatically rejection or misunderstanding.
- Missing calls limit both confirmation and contradiction.
Evidence (2 items)
Supports the claim: New York University; Scott Aaronson and Ernest Davis,
close to zero.
Why it matters here: In problem 14, the correct probability is about 0.289, yet GPT-4 describes the returned fraction as near zero. Problem 13 is another candidate, but its authors explicitly leave the tool expression's correctness unchecked and the extracted expression is incomplete.
- Publisher
- New York University; Scott Aaronson and Ernest Davis
- Kind of source
- Benchmark result
- Describes events of
- Public from
Context only: New York University; Ernest Davis and Scott Aaronson,
error.
Why it matters here: The integral example does not cleanly establish rejection of a correct answer: CI returned an error, and the authors caution that WA's infinity output did not establish divergence. Problem 14's separate unit-conversion failure also does not alone establish misunderstanding or rejection.
- Publisher
- New York University; Ernest Davis and Scott Aaronson
- Kind of source
- Benchmark result
- Describes events of
- Public from
What would settle it
Verify the complete original tool output for the prime-probability example, or locate another distinct transcript showing a correct setup and correct result followed by misinterpretation or rejection.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Getting GPT to work with external tools is harder than you think
In a few cases GPT-4 set up the calculations correctly and the plug-in returned the correct answer, but then GPT-4 either misunderstood or refused to accept that answer.
- Where
- paragraph 15
- Published
- 2023-10-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-getting-gpt-to-work-with-external-0a7866-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.
About assessment marcus-getting-gpt-to-work-with-external-0a7866-a1.