“The field of AI is littered with Band-Aid solutions that work for one problem, and not the next.”
Our reading, the sentence that was checked:
The AI field is full of Band-Aid solutions that work for one problem but not the next.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Scott Alexander’s Misleading Victory Lap, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-scott-alexanders-misleading-victory-776723-a1
We recorded 8 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-07-09. 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
- The field of AI
- littered with Band-Aid solutions
- work for one problem, and not the next
Our interpretation
- The ordinary reading is that narrowly effective methods that fail on related problems recur across AI research. It does not say every method fails to generalize.
- I interpret 'Band-Aid solutions' as methods or interventions whose demonstrated benefits do not reliably transfer to related cases.
- The statement concerns the accumulated research record through publication, rather than a particular reporting quarter.
Counts as supported if: Multiple independent evaluations across different task domains document methods or targeted interventions that succeed on their original tasks but fail on related cases.
Counts as contradicted if: Broad comparative evidence shows that the alleged pattern is exceptional or mistaken, with the relevant methods consistently transferring their gains to related problems.
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
Primary studies in vision and language document the pattern Marcus describes: success on an original test or targeted intervention can leave related problems unsolved. The evidence supports the ordinary qualitative reading of a recurring problem. Both studies also find some successful transfer. They do not support a claim that all AI progress consists of patches or that GPT-4o's particular improvements came from memorizing the bet's prompts.
Limitations
- The result depends on interpreting the metaphor as limited transfer, rather than deliberate engineering of superficial fixes.
- The evidence does not measure field-wide prevalence.
Evidence (2 items)
Supports the claim: arXiv; Rohan Taori and colleagues,
most current techniques provide no robustness
Why it matters here: The study evaluates 204 ImageNet models across 213 conditions and finds limited transfer from synthetic robustness tests to natural changes. It also identifies an exception: larger, more diverse training datasets sometimes improve transfer.
- Where
- Abstract
- Publisher
- arXiv; Rohan Taori and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; R. Thomas McCoy, Ellie Pavlick, and Tal Linzen,
near-zero performance across all three heuristics.
Why it matters here: Models successful on MNLI fail on related inference examples, and targeted augmentation still leaves some withheld categories unsolved. The authors also report successful transfer to other categories and an external dataset, limiting any stronger claim that targeted training never generalizes.
- Publisher
- arXiv; R. Thomas McCoy, Ellie Pavlick, and Tal Linzen
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Scott Alexander’s Misleading Victory Lap
The field of AI is littered with Band-Aid solutions that work for one problem, and not the next.
- Where
- paragraph 10
- Published
- 2025-07-09
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-scott-alexanders-misleading-victory-776723-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-scott-alexanders-misleading-victory-776723-a1.