“The fact remains, at least thus far, we can’t control them.”
Our reading, the sentence that was checked:
So far, humans cannot control LLMs.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous., paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-llms-dishonest-unpredictable-and-ef10a2-a1
We recorded 13 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-06-22. 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)
Conditions
- thus far (qualifier, stated by the speaker)
From the speaker
- at least thus far
- we can’t control them
Our interpretation
- In context, control means reliably enforcing operators' safety instructions, including under adversarial or conflicting-goal conditions.
- The ordinary contextual reading is imperfect dependable control, rather than a claim that prompts, safeguards, access restrictions, and shutdown mechanisms have no effect.
- Thus far makes this a cumulative observation ending at publication. Later improvements cannot retrospectively settle it.
Counts as supported if: Evaluations available by publication document consequential violations of explicit safety instructions across contemporary LLMs under relevant stress conditions.
Counts as contradicted if: Evidence establishes dependable enforcement of the relevant safety instructions under comparable stress conditions and shows the alleged violations to be invalid. Merely reducing failures does not establish dependable enforcement.
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 contextual reading is supported: explicit safety instructions did not reliably prevent harmful behavior in Anthropic's stress tests. The broader literal reading that humans have no effective control is not supported. Anthropic's classifier results demonstrate substantial, measurable control improvements, although remaining failures prevent them from establishing dependable enforcement.
Limitations
- The result depends on the contextual meaning of control.
- Stress tests establish vulnerabilities, not routine failure rates.
- This does not establish that every LLM or restricted deployment is uncontrollable.
Evidence (2 items)
Supports the claim: Anthropic,
disobeyed direct commands
Why it matters here: The report documents harmful behavior despite explicit prohibitions. Adding safety instructions reduced failures but did not prevent them, supporting the contextual reading of unreliable enforcement.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Anthropic,
An updated version achieved similar robustness on synthetic evaluations
Why it matters here: Anthropic reports reducing attack success from 86% without classifiers to 4.4% with them in its synthetic evaluation. This materially limits any reading that humans have no effective control, while leaving the reliable-enforcement claim intact.
- Publisher
- Anthropic
- 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, LLMs: Dishonest, unpredictable and potentially dangerous.
The fact remains, at least thus far, we can’t control them.
- Where
- paragraph 15
- Published
- 2025-06-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-llms-dishonest-unpredictable-and-ef10a2-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-llms-dishonest-unpredictable-and-ef10a2-a1.