““Current deep-learning systems frequently succumb to stupid errors.” Still true.”
Our reading, the sentence that was checked:
Current deep-learning systems still often make stupid errors.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Two years later, deep learning is still faced with the same fundamental challenges, paragraph 16.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-two-years-later-deep-learning-is-6d3049-a1
We recorded 15 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 2024-03-10. 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)
From the speaker
- Current deep-learning systems
- frequently succumb to stupid errors
- Still true
Our interpretation
- The plain reading is that leading systems repeatedly make readily identifiable mistakes on ordinary questions or elementary reasoning tasks. It does not mean every system fails, or that most everyday answers are wrong.
- For this ongoing observation, I use the latest calendar quarter to date before publication: January 1 through March 10, 2024. Research has no standard quarterly reporting cycle; earlier measurements are context only.
- Frequent means substantial recurrence in systematic testing, without imposing a numerical threshold Marcus did not supply.
Counts as supported if: Contemporaneous systematic testing of several leading systems finds substantial rates of clear mistakes on ordinary factual questions or elementary tasks.
Counts as contradicted if: Broad, representative testing finds these mistakes consistently rare across leading systems and ordinary tasks, with apparent examples explained by misleading tests.
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
January 2024 testing found repeated, readily identifiable factual mistakes across five leading systems. Half of the evaluated answers were rated inaccurate. This supports the ordinary reading that clear errors remained a recurring problem. It does not establish that most everyday interactions fail: the sample was small and domain-specific, and GPT-4 performed substantially better than its peers.
Limitations
- The finding depends on reading frequent as recurring in ordinary-task evaluations, rather than as a population-wide frequency threshold.
- Language-model results do not describe every kind of deep-learning system.
- GPT-4's better performance and the study's sampling limitations are material qualifications.
Evidence (2 items)
Supports the claim: Proof News and the Institute for Advanced Study, AI Democracy Projects,
Team majorities rated half of the 130 responses as inaccurate.
Why it matters here: Experts evaluated 26 ordinary election questions across five leading models; errors included incorrect contact details and fabricated information. GPT-4 performed substantially better, with 19% of its responses rated inaccurate, and the authors explicitly warn that this small, nonrandom sample is not representative of everyday usage.
- Where
- Data and Findings
- Publisher
- Proof News and the Institute for Advanced Study, AI Democracy Projects
- Kind of source
- Independent measurement
- Describes events of
- Public from
Context only: Google Research,
the best model achieving 52.9%
Why it matters here: The researchers found difficulty locating simple, unambiguous reasoning mistakes, but the underlying study was already public in 2023. Its deliberately error-enriched dataset also cannot estimate the frequency of mistakes in ordinary model use.
- Publisher
- Google Research
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Two years later, deep learning is still faced with the same fundamental challenges
“Current deep-learning systems frequently succumb to stupid errors.” Still true.
- Where
- paragraph 16
- Published
- 2024-03-10
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-two-years-later-deep-learning-is-6d3049-a1: 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-two-years-later-deep-learning-is-6d3049-a1.