“Outliers remain a problem, too.”
Our reading, the sentence that was checked:
Outliers remain a problem for GPT-4.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, GPT-4’s successes, and GPT-4’s failures, paragraph 4.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-gpt-4s-successes-and-gpt-4s-failures-ef6e9f-a3
We recorded 13 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 2023-03-15. 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
- Outliers remain a problem, too.
- GPT-4
Our interpretation
- The plain reading is that unusual or adversarial inputs still expose failures in GPT-4. This does not assert that every unusual input fails.
- The target is the March 2023 release state, through the post date. I use March month-to-date as the latest public reporting window.
- A formal statistical definition of out-of-distribution performance would require a specified reference distribution, which Marcus does not provide.
Counts as supported if: Release-era evidence documents failures on unusual or adversarial inputs.
Counts as contradicted if: Release-era evaluations establish reliable handling of the relevant unusual inputs and show that the alleged failure cases do not apply to GPT-4.
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 March 14, 2023 disclosure stated that jailbreaks could still make GPT-4 generate content violating its usage guidelines [e1]. This supports the claim's ordinary reading that adversarial inputs could still expose failures at release [e1]. OpenAI also reported improved safety relative to GPT-3.5, alongside those remaining failures [e1].
Limitations
- The disclosure does not measure the prevalence of remaining jailbreak failures or performance on a formally defined out-of-distribution benchmark [e1].
- Adversarial safeguard failures do not establish failures across every class of unusual input [e1].
- The packet provides OpenAI's disclosure without independent replication.
Evidence (1 item)
Supports the claim: OpenAI,
there still exist “jailbreaks”
Why it matters here: OpenAI acknowledges that specially constructed prompts can still defeat GPT-4's safeguards. This supports the ordinary reading that unusual inputs remain a source of failures.
- 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, GPT-4’s successes, and GPT-4’s failures
Outliers remain a problem, too.
- Where
- paragraph 4
- Published
- 2023-03-15
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-gpt-4s-successes-and-gpt-4s-failures-ef6e9f-a3: 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-gpt-4s-successes-and-gpt-4s-failures-ef6e9f-a3.
HeldBy this site
The final-text check failed: citations_correct.
About assessment marcus-gpt-4s-successes-and-gpt-4s-failures-ef6e9f-a2, which was not published.