“In some ways we have made progress, but in the ones I keep harping on—factuality and compositionality—we have not.”
Our reading, the sentence that was checked:
Generative AI has made no progress on factuality and compositionality since 2022, despite progress in other respects.
Assessment: Contradicted Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Horse rides Astronaut, redux, paragraph 25.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-horse-rides-astronaut-redux-f5ef21-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 2024-02-13. 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)
Conditions
- since 2022 (qualifier, stated by the speaker)
From the speaker
- In some ways we have made progress
- factuality and compositionality
- we have not
Our interpretation
- The surrounding retrospective supplies May 2022 as the starting point and February 13, 2024 as the endpoint.
- The ordinary reading is no measurable improvement in factual accuracy or in combining requested objects, attributes, and relations correctly.
- Progress does not require eliminating hallucinations or solving every compositional task.
- Comparable evaluations showing meaningful gains within this interval can contradict the categorical claim; they need not establish improvement on every task.
- The broad wording covers generative AI, not only the horse-and-astronaut prompt.
Counts as supported if: Broad comparable evaluations show no meaningful gains in either factual accuracy or compositional performance over the period.
Counts as contradicted if: Credible comparable evaluations document meaningful improvements in factuality or compositionality during the period.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Contradicted
The dated evidence goes against 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
On the ordinary reading of no measurable progress, contemporary evaluations contradict the claim. GPT-4's report documents improved TruthfulQA performance, and GenEval finds compositional gains across successive image models. These results coexist with serious failures: GenEval still finds spatial relations and attribute binding difficult, and NewsGuard reports persistent misinformation problems. A narrower claim that the problems remained unsolved would be compatible with the evidence.
Limitations
- The verdict depends on reading progress as measurable improvement rather than complete resolution.
- The factuality comparison is company-reported; OpenAI did not check its RLHF data for TruthfulQA contamination.
- Benchmarks cover selected tasks and do not establish reliable performance in every setting.
Evidence (3 items)
Goes against the claim: OpenAI via arXiv,
significantly outperforms
Why it matters here: The report compares factual accuracy on TruthfulQA and reports gains over earlier models. This contradicts a categorical absence of factuality progress, subject to the company's evaluation methods and contamination caveat.
- Publisher
- OpenAI via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt via arXiv,
a significant improvement
Why it matters here: The same benchmark reports higher compositional scores for newer models: overall scores rise from 0.43 for SDv1.5 to 0.55 for SD-XL and 0.61 for IF-XL. Position and attribute binding also improve while remaining difficult.
- Publisher
- Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: NewsGuard,
98 out of the 100
Why it matters here: This audit documents severe persistent misinformation failures and reports no improvement over six months on its narrower test. It limits any claim of universal progress, but does not erase gains on other factuality or compositional evaluations.
- Publisher
- NewsGuard
- Kind of source
- Independent measurement
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Horse rides Astronaut, redux
In some ways we have made progress, but in the ones I keep harping on—factuality and compositionality—we have not.
- Where
- paragraph 25
- Published
- 2024-02-13
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-horse-rides-astronaut-redux-f5ef21-a1: Current. Outcome: Contradicted. 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-horse-rides-astronaut-redux-f5ef21-a1.