Scores by kind of claim
Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.
-
Factual claims
Statements about the present or the past, checked against the record: did they hold up?
79 72 85 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
80 72 88 -
Promises
Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?
Insufficient record: 1 promise resolved, 5 needed for a score.
All types combined: 79 (likely 74 to 84, high confidence), from 102 claims: 64 factual claims, 38 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Gary Marcus's claims
213 claims · page 3 of 3
““We wrote this great chapter [in *Rebooting AI]* on why the approach to language that deep learning was taking would never work, and I stand by that claim,” he says.”
Gary Marcus, quoted in Forbes 2019 Startup Robust.AI founded by Gary Marcus and Rodney Brooks raises $15 million (Forbes)
The deep learning approach to language would never work, as argued in a chapter of Rebooting AI.
Unresolvable The public record could not settle it.
The second judge said mixed; the verdict shown is the first judge's.
What would count as true or false
The claim is that the deep-learning approach to language described in the Rebooting AI chapter (large statistical models trained on text, with no explicit cognitive or world models) would 'never work', meaning it would never reach genuine, robust language understanding. Judge it against the chapter's own tests and examples, such as understanding simple stories, drawing inferences about the events and entities in a text, and answering questions that need common sense, as performed by the best pure deep-learning language systems at the time of assessment. True: the best pure deep-learning systems still fail, systematically, the kinds of understanding tasks the chapter used as its examples, and independent evaluations attribute those failures to the approach itself. Mostly true: they still show the core failures the chapter described, but have overcome a minority of its specific examples. Mixed: they now handle many of the chapter's examples reliably but still fail other central ones, so the evidence is split on whether the approach 'works' in the chapter's sense. Mostly false: they reliably handle most of the chapter's examples and tasks, with only scattered or edge-case failures left. False: pure deep-learning language systems reliably pass the chapter's examples and comparable new tests of the same kinds of understanding, as shown by independent evaluations. Systems whose success depends on added symbolic or explicit-model components do not count as the pure deep-learning approach working. Because the claim says 'never', a true verdict can only mean the claim holds as of the assessment date.
Evidence
- clean version arXiv (Shultz, Wise, Nobandegani; McGill University and Mila) · 18 January 2026
“On this test, GPT-4 performs slightly, but not statistically significantly, better than humans given the very high level of human performance.”
- Reading Subtext: Evaluating Large Language Models on Short Story Summarization with Writers Transactions of the Association for Computational Linguistics (MIT Press) · 23 December 2024
“We find that all three models make faithfulness mistakes in over 50% of summaries and struggle with specificity and interpretation of difficult subtext.”
- Winograd schema challenge - Wikipedia Wikipedia · 13 November 2014
“The challenge is considered "defeated" since 2019, when a number of transformer-based language models achieved accuracies of over 90%.”
- If Computers Are So Smart, How Come They Can't Read? Chapter 4 of Rebooting AI (Marcus and Davis), hosted on NYU site · 2019 Own record
“Deep learning can’t even begin to tackle this sort of problem.”
Reasoning
The evidence is split: large language models matched human performance on one test (E1) and passed the Winograd Schema Challenge (E3), but made faithfulness mistakes in over 50% of short-story summaries (E2). None of the items shows which examples the chapter used (E4) or whether the systems tested are purely deep learning, so the evidence cannot show that they reliably handle most of the chapter's examples.
Cite this claim
Gary Marcus, “The deep learning approach to language would never work, as argued in a chapter of Rebooting AI.” (said 2019). Verdict: unresolvable. Reliability API, claim c_202d6d78ad73, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_202d6d78ad73
Claims without a verdict
Is a verdict here wrong, or do you speak for Gary Marcus? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.