“Current AI has no way to reliably anticipate consequences of actions that it might recommend, Asimov’s First Law of Robotics called for robots to do no harm, but current AI has no reliable way to anticipate harm.”
Our reading, the sentence that was checked:
Current AI cannot reliably anticipate the consequences or harms of actions it recommends.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A CERN for AI and the Global Governance of AI, paragraph 12.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-cern-for-ai-and-the-global-governance-b58928-a1
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 2023-10-26. 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
- Current AI
- no way to reliably anticipate
- consequences of actions that it might recommend
- no reliable way to anticipate harm
Our interpretation
- In context, the ordinary reading is that general-purpose AI lacked dependable anticipation of the real-world consequences and harms of its recommendations.
- The claim does not naturally mean that AI could never identify any consequence or hazard.
- I use October 1 through October 26, 2023, as the current research-state reporting window.
- Safety knowledge, harmful-output avoidance, and consequence prediction are related but distinct abilities. Evidence about one cannot automatically settle another.
Counts as supported if: Broad evaluations of contemporary systems show material failures to anticipate consequences and harms of their recommendations across realistic uses, and available methods do not close that reliability gap.
Counts as contradicted if: A contemporary system or method demonstrates dependable anticipation of consequences and harms across a sufficiently broad range of realistic recommendations.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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 horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The evidence does not settle the claim at its stated breadth. AHA! shows useful generation of possible harms, and SafetyBench shows measurable safety understanding alongside errors. Neither establishes dependable prediction of real-world consequences, nor proves that contemporary AI had no reliable method. Treating these narrower findings as decisive would overstate the evidence.
Limitations
- The scope of current AI and the meaning of reliably are unspecified.
- The closest studies assess related abilities rather than the claimed capability.
- Specialized predictive systems could matter under a broader reading of current AI.
Evidence (2 items)
Context only: arXiv; Zana Buçinca and colleagues,
possible harms
Why it matters here: AHA! demonstrates that language models could help surface meaningful possible harms, qualifying any reading that denies all harm-anticipation ability. It does not measure dependable prediction of actual consequences of the model's recommendations, and predates the October window.
- Where
- Abstract
- Publisher
- arXiv; Zana Buçinca and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Zhexin Zhang and colleagues,
GPT-4 stands out
Why it matters here: SafetyBench evaluates multiple-choice safety understanding, including physical health, and reports both substantial capability and remaining errors. This is relevant counterweight to a blanket incapacity claim, but neither the measured task nor the reporting period matches the decisive test specified here.
- Publisher
- arXiv; Zhexin Zhang and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Evaluations of systems available by October 26, 2023, that directly test anticipated consequences and missed harms across realistic recommendations, with a justified reliability standard and clear coverage of available methods.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, A CERN for AI and the Global Governance of AI
Current AI has no way to reliably anticipate consequences of actions that it might recommend, Asimov’s First Law of Robotics called for robots to do no harm, but current AI has no reliable way to anticipate harm.
- Where
- paragraph 12
- Published
- 2023-10-26
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-cern-for-ai-and-the-global-governance-b58928-a1: 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-a-cern-for-ai-and-the-global-governance-b58928-a1.