“Yejin Choi worried that LLMs were not reliable. They still aren’t.”
Our reading, the sentence that was checked:
LLMs remained unreliable as of April 2024, a year after Choi's TED talk.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, TED, AI, and the test of time, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-ted-ai-and-the-test-of-time-a42209-a1
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 2024-04-15. 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
- LLMs were not reliable
- They still aren’t
Our interpretation
- The plain reading is that general-purpose LLMs continued to make recurring factual or reasoning errors that prevented users from depending on their answers without checking.
- This does not mean that every answer was wrong, that no application was reliable, or that models had stopped improving.
- Use the most recent completed quarter before publication, January through March 2024, for contemporaneous evaluations. Marcus supplied no reporting period or numerical reliability threshold.
Counts as supported if: Contemporaneous evaluations across multiple prominent models document recurring factual or reasoning errors in substantive tasks.
Counts as contradicted if: Broad contemporaneous evaluations establish dependable factual and reasoning performance across the claimed scope, with failures confined to exceptional cases.
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
Contemporaneous evaluations document recurring factual errors across multiple model families and tasks. That supports the ordinary reading that users still needed to check LLM answers. The evidence also shows improvements; Marcus’s statement does not require accuracy to have remained unchanged.
Limitations
- The finding concerns general factual dependability, not every task or application.
- The evidence uses task-specific evaluations rather than a representative sample of all user requests.
- Evaluation months identify when the findings were reported; exact test-run dates are not supplied in the cited passages.
Evidence (2 items)
Supports the claim: arXiv; Jerry Wei and colleagues,
contains factual errors
Why it matters here: The March evaluation covers thirteen models from the Gemini, GPT, Claude and PaLM-2 families and documents continuing factuality limitations. It also finds better performance among larger models, so it supports continued unreliability without supporting stagnation.
- Publisher
- arXiv; Jerry Wei and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Liyan Tang and colleagues,
factual errors
Why it matters here: Human annotations identify recurring factual errors in topic-focused dialogue summaries. This supplies evidence from another substantive task, although it does not establish an error rate for all uses.
- Publisher
- arXiv; Liyan Tang and colleagues
- 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, TED, AI, and the test of time
Yejin Choi worried that LLMs were not reliable. They still aren’t.
- Where
- paragraph 5
- Published
- 2024-04-15
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-ted-ai-and-the-test-of-time-a42209-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-ted-ai-and-the-test-of-time-a42209-a1.