“More ominously, and less amusingly, a new test by Anthropic shows that under the right circumstances, an LLM might be prepared to kill a human.”
Our reading:
Anthropic's new test shows that under the right circumstances an LLM might be prepared to kill a human.
Assessment: Can’t be checked
What happened
Not graded on evidence
Outcome: Can’t be checked
An opinion, a recommendation, or too vague to test. Listed, not scored.
Evidence up to . Checked .
Why
Marcus says an LLM 'might' be prepared to kill under unspecified circumstances. Anthropic's report provides a factual basis for that warning through a simulated lethal-action scenario. But the supplied review excludes scoring possibility language. Calling this supported would require replacing his qualified statement with the narrower claim that a model selected a lethal action in a simulation. The simulation caveat also does not contradict what he actually said.
Caveats
- This disposition concerns eligibility for scoring, not whether the warning is reasonable.
- The report establishes neither an actual killing nor the frequency of such behavior in real deployments.
Evidence (1 item)
Claims like this aren’t scored, so there is no evidence to show.
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- under the right circumstances (their words)
- per Anthropic test (their words)
Where we had to interpret
- The statement expresses a possibility inferred from a test, rather than a firm prediction or an assertion that a person was killed.
- The surrounding text rejects attributing human desires to models. I therefore read 'prepared' as referring to potentially harmful outputs or actions, not a conscious intention.
- Whether Anthropic reported a simulated lethal-action scenario is separately checkable. Substituting that narrower observation for the possibility-qualified proposition would change what is being graded.
- Following the supplied reviewer objection and the rule for hedged claims, I do not convert 'might' into a scored existential claim.
We don’t grade claims like this. We only check the quote and who said it.
What would settle it
A separately submitted claim about the specific behavior observed in Anthropic's simulation could be assessed. It should not be substituted for this wording.
We found 18 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous., paragraph 8.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2025-06-22.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous.
More ominously, and less amusingly, a new test by Anthropic shows that under the right circumstances, an LLM might be prepared to kill a human.
- Where
- paragraph 8
- Published
- 2025-06-22
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Can’t be checked. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.