“The problem for the neural network community is not that I have changed my goal posts, it’s that my 2012 concerns still stand.”
Our reading, the sentence that was checked:
Marcus's 2012 concerns about deep learning still stand, and the problem for the neural network community is not that he has changed his goal posts.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The New Sport of Misrepresenting AI Criticism, paragraph 23.
- 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-50b5c8-a1
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 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
- not that I have changed my goal posts
- my 2012 concerns still stand
- representing causal relationships
- acquiring abstract ideas
- performing logical inferences
- integrating abstract knowledge
Our interpretation
- The plain reading is that the listed problems remained substantial unresolved limitations in 2022, and that invoking them continued his earlier criteria.
- This does not require that neural networks had made no progress or could never perform these tasks.
- The target period spans the claimed continuity from 2012 through publication. Evidence of remaining technical limitations must describe systems evaluated in 2022.
- Support for one technical concern does not automatically establish all four concerns or the separate assertion about unchanged criteria.
Counts as supported if: Contemporary evaluations establish substantial remaining limitations across the listed capabilities, and dated records establish continuity of the relevant criteria from 2012 to September 2022.
Counts as contradicted if: Contemporary evaluations establish robust resolution of the listed concerns, or dated records show a material change in the relevant success criteria that contradicts the claimed continuity.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
The cited research supports a specific remaining weakness: learned logical performance did not generalize reliably across distributions. It also demonstrates that BERT can represent a correct reasoning procedure. Neither result settles the status of all the listed concerns or proves that Marcus's relevant criteria remained unchanged. The complete claim therefore remains unresolved.
Limitations
- Evidence about one model and logical task cannot settle every capability of deep learning.
- Persistent practical weaknesses and architectural impossibility are different claims.
Evidence (2 items)
Supports the claim: arXiv; Honghua Zhang and colleagues,
generalize
Why it matters here: The experiments show that trained BERT models could perform well on familiar distributions while failing changed distributions within the same logical problem space. This supports the continued concern about robust learned inference, but not the entire compound claim.
- Publisher
- arXiv; Honghua Zhang and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Honghua Zhang and colleagues,
capacity
Why it matters here: The same paper constructs parameters that implement logical reasoning. This limits any categorical reading that neural architectures lack a way to perform inference, while leaving the observed training and generalization problem intact.
- Publisher
- arXiv; Honghua Zhang and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A documented comparison of the relevant historical criteria, together with independent 2022 evaluations covering causal representation, abstract relations, logical inference, and object-function knowledge.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, The New Sport of Misrepresenting AI Criticism
The problem for the neural network community is not that I have changed my goal posts, it’s that my 2012 concerns still stand.
- Where
- paragraph 23
- Published
- 2022-09-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-new-sport-of-misrepresenting-50b5c8-a1: 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-the-new-sport-of-misrepresenting-50b5c8-a1.