“The shocking (yet, in historical context, see below, unsurprising) answer is no:”
Our reading, the sentence that was checked:
LLMs trained on 'A is B' cannot automatically infer that 'B is A'.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Elegant and powerful new result that seriously undermines large language models, paragraph 3.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-elegant-and-powerful-new-result-that-99694d-a2
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 2023-09-22. 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
- trained on statements of the form A is B (qualifier, stated by the speaker)
From the speaker
- LLMs trained on A is B
- infer automatically that B is A
- answer is no
Our interpretation
- Read in context as a description of the cited experiments, not a proof about every possible language model.
- Automatic inference means answering the inverse question from facts learned during training, without putting the original fact in the prompt.
- The target is the experiments reported in 2023 before the post. Their exact execution dates are not specified.
Counts as supported if: Controlled tests show successful learning in the trained direction but reverse answers or answer probabilities no better than chance.
Counts as contradicted if: The cited tests show reliable transfer to the reverse direction without reverse-direction training or supplying the fact in the prompt.
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
The September 21, 2023 preprint reports that fine-tuned GPT-3-175B achieved 50.0% and 96.7% accuracy in the trained directions, versus 0.0% and 0.1% in the reverse directions [e1, Table 1]. It also found no detectable probability advantage for the correct reversed name over random names [e1, section 2.1.2]. This supports failure of automatic reversal in the tested training setup; the paper explicitly excludes inference from facts supplied in the prompt from this failure [e1, section 1].
Limitations
- These experiments do not establish impossibility across all models, training procedures or prompting methods [e1].
- The real-celebrity experiment is tentative because the models' precise training data were unknown [e1].
- The preprint's publication date is known, but exact experiment execution dates are not supplied [e1].
Evidence (1 item)
Supports the claim: arXiv; Berglund et al.,
Results
Why it matters here: Table 1 reports strong directional asymmetry, and section 2.1.2 finds no detectable advantage for the correct reversed name over random names. Section 1 explicitly excludes in-context learning from this failure.
- Publisher
- arXiv; Berglund et al.
- 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, Elegant and powerful new result that seriously undermines large language models
The shocking (yet, in historical context, see below, unsurprising) answer is no:
- Where
- paragraph 3
- Published
- 2023-09-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-elegant-and-powerful-new-result-that-99694d-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-elegant-and-powerful-new-result-that-99694d-a2.