“It has become apparent that large language models, OpenAI’s core technology, sometimes leak private data, and that this problem has not been resolved; hallucinations too remain problematic.”
Our reading, the sentence that was checked:
LLMs sometimes leak private data, an unresolved problem, and hallucinations remain problematic.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, OpenAI’s Got 9.9 Problems, and Twitch Ain’t One, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-openais-got-99-problems-and-twitch-0200e7-a1
We recorded 25 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 2024-01-27. 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
- large language models
- sometimes leak private data
- this problem has not been resolved
- hallucinations too remain problematic
Our interpretation
- The claim concerns persistent classes of LLM failure, not every model or every interaction.
- Private data is read as personal or sensitive information disclosed through model output, including memorized personal information originally collected from public sources. Confidential user conversations are a narrower category.
- Unresolved means the underlying disclosure risk remained, even if a particular extraction technique was blocked.
- Problematic hallucinations means material factual errors in consequential uses, without imposing a universal error-rate threshold.
- For the ongoing observation, use the latest completed quarter and the elapsed publication quarter: October 2023 through January 27, 2024.
Counts as supported if: Contemporaneous research establishes continuing personal-data disclosure vulnerabilities and material hallucination problems, while distinguishing individual mitigations from a general solution.
Counts as contradicted if: Reliable evidence establishes that either asserted ongoing problem had been comprehensively resolved for the relevant LLM class during the target period.
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 research supports the ordinary, class-level reading: LLMs had disclosed verified personal information, researchers identified continuing underlying vulnerabilities, and the January legal evaluation found serious hallucination problems. A reported patch to the particular word-repetition attack limits the finding but does not establish a general solution. Support depends on reading private data to include personal information recovered from training data, rather than exclusively confidential conversations.
Limitations
- The extraction experiments preceded the target period; the contemporaneous outcome evidence concerns the continuing vulnerability assessment.
- Personal information recovered from public-source training material is not necessarily secret information.
- Legal benchmark error rates cannot be generalized to every LLM task.
- The patch report is secondary reporting.
Evidence (4 items)
Supports the claim: Authors of Scalable Extraction of Training Data from (Production) Language Models,
language models are subject to divergence
Why it matters here: This is the researchers’ dated assessment of the continuing vulnerability, not a claim that their experiments occurred on publication day. They explain why a word-repetition filter would address one exploit without eliminating memorization and divergence.
- Publisher
- Authors of Scalable Extraction of Training Data from (Production) Language Models
- Kind of source
- Demonstration
- Describes events of
- Public from
Supports the claim: Dahl, Magesh, Suzgun and Ho, arXiv,
unsupervised integration of popular LLMs
Why it matters here: The January assessment reports legal hallucinations ranging from 69% for ChatGPT 3.5 to 88% for Llama 2 on the specified questions. The date identifies the published assessment of the problem; the abstract does not specify individual test dates.
- Publisher
- Dahl, Magesh, Suzgun and Ho, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Nasr and coauthors, arXiv,
We recover personally identifiable information of dozens of individuals.
Why it matters here: The researchers verified extracted personal information against their comparison dataset. The attack was discovered and disclosed before the target period, so this is background evidence rather than a newly dated extraction outcome.
- Publisher
- Nasr and coauthors, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: SecurityWeek,
OpenAI has been notified and the attack no longer works.
Why it matters here: This materially limits any assertion that the original ChatGPT exploit remained usable. The same report distinguishes that mitigation from resolving the underlying vulnerabilities.
- Publisher
- SecurityWeek
- Kind of source
- News report
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, OpenAI’s Got 9.9 Problems, and Twitch Ain’t One
It has become apparent that large language models, OpenAI’s core technology, sometimes leak private data, and that this problem has not been resolved; hallucinations too remain problematic.
- Where
- paragraph 20
- Published
- 2024-01-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-openais-got-99-problems-and-twitch-0200e7-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-openais-got-99-problems-and-twitch-0200e7-a1.