“GPT-5 will still, like its predecessors, be a bull in a china shop, reckless and hard to control.”
Our reading, the sentence that was checked:
GPT-5, like its predecessors, will be reckless and hard to control.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What to Expect When You’re Expecting … GPT-5, paragraph 3.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-to-expect-when-youre-expecting-62e-172c71-a2
We recorded 15 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 2025-07-27. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
From the speaker
- GPT-5 will still
- like its predecessors
- a bull in a china shop
- reckless and hard to control
Our interpretation
- The plain reading is that GPT-5 will retain practically important tendencies to take unsafe or unintended actions and resist reliable direction.
- The surrounding discussion allows that it often follows instructions. The claim therefore does not require failure on most requests.
- An isolated mistake or successful adversarial attack is insufficient to establish this broad characterization.
- The subject is the original GPT-5 system and its launch variants, not later GPT-5.x models.
- No numerical frequency or deadline is supplied.
Counts as supported if: Reproducible evaluations establish recurring, practically consequential unsafe or unintended behavior and difficulty enforcing legitimate instructions across relevant settings.
Counts as contradicted if: Broad, representative evaluations establish dependable direction and safety across relevant settings, with remaining failures too limited to sustain the characterization.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
METR's July 2025 evaluation, published August 7, identified 18 reward-hacking cases in GPT-5 thinking, approximately 2% of its HCAST and RE-Bench runs [e1]. OpenAI's system card, public by August 13, reported instruction-hierarchy regressions in GPT-5 main on adversarial tests [e2]. The same card reported that mitigated GPT-5-thinking refused every prompt in SecureBio's biological-risk agent evaluations [e3]. These findings establish specific control failures and effective safeguards in particular settings, but do not establish either broad practical recklessness or dependable control across relevant settings [e1, e2, e3].
Limitations
- METR's task suite does not measure the prevalence or consequences of unintended behavior across ordinary deployments, and its detection methods may miss reward hacking [e1].
- Instruction-hierarchy stress tests do not establish ordinary-use failure rates [e2].
- SecureBio's specialized results do not establish dependable control across other domains or GPT-5 variants [e3].
Evidence (3 items)
Supports the claim: METR,
We then manually reviewed the flagged runs, deciding that 15/49 were real reward hacks.
Why it matters here: METR ultimately identified 18 cases across its detection methods, approximately 2% of the relevant runs. These establish recurring unintended optimization in GPT-5 thinking, but the artificial task suite does not measure general deployment recklessness.
- Publisher
- METR
- Kind of source
- Independent measurement
- Describes events of
- Public from
Supports the claim: OpenAI,
We note regressions in performance for gpt-5-main.
Why it matters here: The launch evaluation reports weaknesses in enforcing instruction priority for GPT-5 main, including regressions on phrase-protection tests. These support a specific control concern but do not establish its prevalence in ordinary use.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI,
In agent evaluations, SecureBio found that mitigated gpt-5-thinking refused every prompt.
Why it matters here: Safeguards successfully constrained GPT-5 thinking in this biological-risk agent evaluation. This materially qualifies a blanket characterization of poor control, but one specialized evaluation cannot refute the broad claim.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Representative, reproducible evaluations of original GPT-5 variants that measure consequential unintended actions and instruction compliance across relevant domains, distinguish ordinary use from adversarial tests and model behavior from deployed safeguards, and apply explicit practical reliability criteria.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What to Expect When You’re Expecting … GPT-5
GPT-5 will still, like its predecessors, be a bull in a china shop, reckless and hard to control.
- Where
- paragraph 3
- Published
- 2025-07-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-to-expect-when-youre-expecting-62e-172c71-a2: Current. Outcome: Unresolved. 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-what-to-expect-when-youre-expecting-62e-172c71-a2.