“That’s still true – and it very much applies in medicine.”
Our reading, the sentence that was checked:
Large language models are still frequently wrong yet confident, producing authoritative bullshit that shouldn't be trusted, and this applies strongly in medicine.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Please don’t trust your chatbot for medical advice, paragraph 2.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-please-dont-trust-your-chatbot-for-3db366-a1
We recorded 6 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 2026-04-21. 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
- in medicine (qualifier, stated by the speaker)
From the speaker
- frequently wrong, never in doubt
- That’s still true
- very much applies in medicine
- about one year ago
Our interpretation
- The testable core is that commonly used LLMs repeatedly give materially incorrect medical answers with unjustified confidence.
- I read the language as a warning against unverified reliance, not a claim that every answer is wrong or that medical applications have no value.
- The context explicitly invokes approximately year-old testing. I interpret the observation as covering the 2025-to-publication evidence period; extending those findings to every model available on publication day is a separate inference.
Counts as supported if: Independent evaluations across commonly used models document recurring medical errors together with confident presentation, with additional evidence that errors can affect consequential medical decisions.
Counts as contradicted if: Comparable evaluations find medical errors exceptional and show that models reliably communicate uncertainty when wrong.
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
The ordinary warning has empirical support: tested chatbots produced recurring medical errors with confident wording, and a newer health product missed consequential emergencies. This supports caution about unverified reliance. It does not establish that every answer is wrong, every model is equally unreliable, or every medical application lacks value.
Limitations
- Support depends on the contextual evidence-period reading.
- Confidence in the BMJ audit was assessed descriptively, not through a calibrated numerical measure.
- Marcus's surrounding description overstates the highly problematic share: it was 19.6%, not nearly half.
- Stress-test results cannot be treated as population-wide error rates.
Evidence (3 items)
Supports the claim: BMJ Open,
amplifying misinformation.
Why it matters here: Across five chatbots, 49.6% of answers were problematic, including 19.6% highly problematic; researchers observed confident presentation even when answers were incorrect. These were misinformation-oriented stress prompts, not a representative sample of patient conversations.
- Publisher
- BMJ Open
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Supports the claim: Nature Medicine,
51.6% (33/64) were undertriaged
Why it matters here: Testing on January 9–11 documents consequential errors in a newer consumer health system. Performance varied substantially: semi-urgent cases achieved 93.0% accuracy, and a supplementary set of textbook emergencies had no undertriage.
- Publisher
- Nature Medicine
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: JAMA Network Open,
final diagnosis, management
Why it matters here: The 21-model evaluation found substantial differences between reasoning stages, including strong final-diagnosis performance. This limits a blanket reading of medical incompetence but does not negate recurring confident errors.
- Publisher
- JAMA Network Open
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Please don’t trust your chatbot for medical advice
That’s still true – and it very much applies in medicine.
- Where
- paragraph 2
- Published
- 2026-04-21
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-please-dont-trust-your-chatbot-for-3db366-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-please-dont-trust-your-chatbot-for-3db366-a1.