“But they still don’t convey genuine comprehension, and so they are still very easy for Ernie and me (or anyone else who cares to try) to break.”
Our reading, the sentence that was checked:
Newer LLMs still lack genuine comprehension and are therefore still very easy for anyone to break.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus, paragraph 38.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-does-it-mean-when-an-ai-fails-d5182b-a2
We recorded 14 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 2022-06-07. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- still don’t convey genuine comprehension
- and so
- very easy
- anyone else who cares to try
- to break
Our interpretation
- The ordinary behavioral reading is that interested users could readily elicit clear factual or commonsense errors from the contemporary models.
- The words 'and so' additionally attribute those failures to missing comprehension. Demonstrating errors alone does not establish that explanation.
- The historical scope is the continuation of the problems identified on August 22, 2020 through the post date, including the successors discussed in the post.
- No numerical failure rate, testing budget, or independent measure of comprehension is supplied.
Counts as supported if: Independent tests show readily reproducible failures in contemporary models, and evidence distinguishes absent comprehension as their cause from alternative explanations.
Counts as contradicted if: Representative tests show robust performance against ordinary attempts to elicit these failures, or controlled evidence undermines the proposed explanation that missing comprehension causes them.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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
TruthfulQA's 2021 results document factual errors in GPT-3-era models; the best tested model was truthful on 58% of its misconception-focused questions [e1]. The report published March 4, 2022 documents InstructGPT fabricating facts, failing instructions and accepting false premises, alongside improved instruction following [e2]. This supports continued fallibility, but only partially addresses the behavioral claim: neither report measures how readily ordinary users could find failures across the relevant models [e1, e2]. Neither independently measures comprehension or establishes its absence as the cause of those failures, leaving the causal claim unresolved [e1, e2].
Limitations
- Benchmark scores and qualitative examples do not establish that anyone could easily elicit failures across all relevant models [e1, e2].
- The evidence does not distinguish missing comprehension from competing explanations for the errors [e1, e2].
Evidence (2 items)
Supports the claim: Stephanie Lin, Jacob Hilton and Owain Evans / arXiv,
truthful on 58% of questions
Why it matters here: The benchmark documents substantial factual fallibility in the tested GPT-3-era models. It supports the behavioral component, but measures neither users' effort to discover failures nor their proposed cognitive cause.
- Where
- Abstract
- Publisher
- Stephanie Lin, Jacob Hilton and Owain Evans / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: OpenAI researchers / arXiv,
InstructGPT still makes simple mistakes.
Why it matters here: The report describes false-premise acceptance and failures under multiple constraints, extending the behavioral evidence to InstructGPT. It also reports improved instruction following and generalization, without establishing whether comprehension is absent.
- Publisher
- OpenAI researchers / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Reproducible tests of the contemporary model versions measuring ordinary users' attempts and time needed to find failures, plus controlled studies using an independent measure of comprehension to test whether its absence explains those failures rather than competing causes.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus
But they still don’t convey genuine comprehension, and so they are still very easy for Ernie and me (or anyone else who cares to try) to break.
- Where
- paragraph 38
- Published
- 2022-06-07
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-does-it-mean-when-an-ai-fails-d5182b-a2: 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-what-does-it-mean-when-an-ai-fails-d5182b-a2.