“Nowadays, we have “multimodal” systems, trained on both text and images, and the aforementioned “hallucinations” (perhaps better called failed approximations, as Gerben Wierda has pointed out) and other errors persist and even crop up in new ways.”
Our reading, the sentence that was checked:
Multimodal systems trained on text and images still hallucinate and make other errors, sometimes in new ways.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Sora Can’t Handle the Truth, paragraph 9.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-sora-cant-handle-the-truth-a606ad-a2
We recorded 11 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 2024-02-17. 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
- Nowadays
- “multimodal” systems, trained on both text and images
- “hallucinations”
- other errors persist
- crop up in new ways
Our interpretation
- The ordinary reading is that adding visual training has not eliminated fabricated content or errors, and that visual systems exhibit additional forms of error.
- This does not assert that every model fails, that error rates have not improved, or that multimodal training provides no benefit.
- For this ongoing observation, I use the publication month through the post date. Model demonstrations and evaluations have no standard quarterly reporting period.
- New ways means errors involving visual content or consistency, rather than a claim of historically unprecedented error categories.
Counts as supported if: Contemporaneous multimodal outputs or evaluations document fabricated content and visual errors despite training involving text and visual material.
Counts as contradicted if: The cited failures are shown to be misidentified, and reliable evaluations establish that the relevant contemporary multimodal systems no longer exhibit the claimed errors.
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
OpenAI’s February 15, 2024 Sora launch disclosure identifies spontaneously appearing animals or people, physically implausible motion, and unnatural object morphing [e1]. These documented failures support the claim that invented content and visual errors persisted in a contemporary multimodal system during February 1–17, 2024 [e1].
Limitations
- The demonstrations establish that errors occur, not their frequency across multimodal systems or whether multimodal training reduces them.
- The GPT-4V findings published September 25, 2023 provide earlier context, not evidence of tests during February 2024 [e2].
- The evidence documents visual forms of error, not historically unprecedented error categories.
Evidence (2 items)
Supports the claim: OpenAI,
Animals or people can spontaneously appear
Why it matters here: The launch material documents invented entities in generated video. This supports the ordinary reading that multimodal generation retains fabrication errors and exhibits visual consistency failures.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI,
the model exhibited some key limitations.
Why it matters here: The surrounding evaluation documents hallucinations, factual mistakes, and erroneous merging of nearby text in images. These earlier tests provide background, rather than outcome evidence for February 2024.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Sora Can’t Handle the Truth
Nowadays, we have “multimodal” systems, trained on both text and images, and the aforementioned “hallucinations” (perhaps better called failed approximations, as Gerben Wierda has pointed out) and other errors persist and even crop up in new ways.
- Where
- paragraph 9
- Published
- 2024-02-17
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-sora-cant-handle-the-truth-a606ad-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-sora-cant-handle-the-truth-a606ad-a2.