“But the smart money is still on being cautious about this, and in recognizing that what the Facebook researchers said in 2019 is still the case.”
Our reading, the sentence that was checked:
State-of-the-art models still learn to exploit spurious statistical patterns rather than learning meaning flexibly as humans do, so benchmark gains warrant caution.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The New Sport of Misrepresenting AI Criticism, paragraph 44.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-the-new-sport-of-misrepresenting-cfd615-a2
We recorded 4 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 2022-09-23. 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)
From the speaker
- state-of-the-art models learn to exploit spurious statistical patterns in datasets
- instead of learning meaning in the flexible and generalizable way that humans do
- what the Facebook researchers said in 2019 is still the case
Our interpretation
- The checkable assertion is that reliance on misleading statistical cues remained a problem for high-performing NLP systems in 2022. The advice to be cautious is not separately scored.
- The plain reading concerns incomplete, unreliable generalization despite benchmark success. It does not require models to have learned no meaningful linguistic information whatsoever.
- The ongoing state has no named reporting period. I use research reported during 2022 through September 23, the current research year up to publication; this field has no standard quarterly reporting cycle.
- 'State-of-the-art' is read as strong task-performing NLP models discussed in contemporary research, rather than every model or only the largest model released immediately before the post.
Counts as supported if: Contemporary research finds that strong NLP models still rely on task-irrelevant cues or improve on controlled generalization tests when those cues are reduced, despite doing well on established tasks.
Counts as contradicted if: Contemporary controlled studies establish that the relevant strong models have overcome the alleged shortcut dependence and reliably track task meaning when incidental statistical cues change.
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
March 2022 experiments found that reducing misleading dataset correlations improved generalization for BERT, RoBERTa and ALBERT, supporting continued shortcut dependence during the period before September 23, 2022 [e6]. An April 2022 example found that sentiment-fine-tuned RoBERTa-large assigned an ambiguous sentence a positive label with 99.99% confidence; attribution methods highlighted sentiment-associated words [e7]. The March study also found better generalization from larger models and debiasing, so persistent shortcut dependence coexisted with genuine improvements [e6].
Limitations
- The evidence covers selected NLP models and tasks, not every leading model available by September 23, 2022 [e6, e7].
- The single RoBERTa example does not establish prevalence [e7].
- These studies do not establish a total absence of meaningful learning or directly measure the full breadth of human semantic flexibility [e6, e7].
Evidence (2 items)
Supports the claim: arXiv; ACL 2022 author manuscript,
generalise better
Why it matters here: Reducing misleading correlations improved independent-test performance for BERT and stronger RoBERTa and ALBERT models, supporting continued susceptibility in 2022. The same experiments found better generalization from larger models and debiasing, which limits any claim that no genuine progress occurred.
- Publisher
- arXiv; ACL 2022 author manuscript
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Supports the claim: arXiv; Findings of NAACL 2022 author manuscript,
99.99% confidence.
Why it matters here: A sentiment-fine-tuned RoBERTa-large model confidently classified an ambiguous sentence as positive; attribution methods highlighted words associated with positive sentiment. This illustrates persistent cue dependence, although the authors explicitly say the example does not establish prevalence.
- Publisher
- arXiv; Findings of NAACL 2022 author manuscript
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, The New Sport of Misrepresenting AI Criticism
But the smart money is still on being cautious about this, and in recognizing that what the Facebook researchers said in 2019 is still the case.
- Where
- paragraph 44
- Published
- 2022-09-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-new-sport-of-misrepresenting-cfd615-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-the-new-sport-of-misrepresenting-cfd615-a2.