Pundit ClaimsAI claims by Gary Marcus and Ed Zitron, with the evidence

Gary Marcus Marcus on AI Published

Is the US military actually afraid of Claude? A new theory of why Anthropic was labeled a supply chain risk.

We recorded 10 claims from this post. 3 were checked against evidence. 7 can’t be checked. 5 more are held. Read the post at its source.

Claims from this post (10)

In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.

  1. Assessment: SupportedSaid

    “It is a misapprehension of LLMs to confuse the things they say with what is true.”

    What we tested: What LLMs say should not be taken as true just because they say it; treating their statements as truth misunderstands them.

    What happened: Table 2 records eight hallucinations for moonshotai/kimi-k2 on February 1, checked by the biographical subject. The same day's Claude Opus 4.5 result has zero, showing that the finding concerns fallibility rather than universal failure. (Martin Monperrus, )

  2. Assessment: Can’t be checkedSaid

    “They mimic things humans say, but we shouldn’t take them as veridical descriptions of anything happening on the inside.”

    What we tested: LLMs mimic human speech, so their statements should not be treated as accurate descriptions of their internal states.

  3. Assessment: Can’t be checkedSaid

    “Claude is next-word predictor, not an authority on its own internal states. It’s silly to take its estimates seriously.”

    What we tested: Claude is a next-word predictor and not an authority on its own internal states, so it is silly to take its estimates of its own consciousness seriously.

  4. Assessment: SupportedSaid

    “the fact that LLMs hallucinate (patently true)”

    What we tested: LLMs hallucinate.

    What happened: OpenAI reports hallucinations in its release evaluation, including after improvements. The reported reductions imply residual hallucinations; they are not claims of elimination. (OpenAI, )

  5. Assessment: Can’t be checkedSaid

    “If you actually believed that Claude was sentient and that sentience was de facto a supply chain risk, you would have to think that essentially every LLM is a supply chain risk. In which case the right choice is to stop using LLMs altogether, not to selectively punish Anthropic.”

    What we tested: Anyone who actually believed that Claude was sentient and that sentience was de facto a supply chain risk would have to think that essentially every LLM is a supply chain risk, in which case the right choice would be to stop using LLMs altogether rather than selectively punish Anthropic.

  6. Assessment: Can’t be checkedSaid

    “But there is no more reason to think that Claude is sentient than to think that Google’s now long-obsolete system LAMDA was.”

    What we tested: There is no more reason to think Claude is sentient than to think Google's LaMDA was.

  7. Assessment: Can’t be checkedSaid

    “If you are worried about systems having guardrails, you shouldn’t be using LLMs, because you certainly shouldn’t be using them without guardrails.”

    What we tested: LLMs should never be used without guardrails, so anyone worried about guardrails should not use LLMs at all.

  8. Assessment: SupportedSaid

    “Essentially the same argument applies to Claude’s constitution, which amounts to a set of built-in guardrails. OpenAI has those, too!”

    What we tested: Claude's constitution is basically a set of built-in guardrails, and OpenAI's models have guardrails too.

    What happened: Anthropic directly connects the constitution to model training. The same announcement describes safety priorities and hard behavioral constraints, supporting the broad guardrails reading. (Anthropic, )

  9. Assessment: Can’t be checkedSaid

    “If you think it is bad to use hallucinating LLMs to autonomously choose targets (I certainly do), you shoudn’t use any of them.”

    What we tested: No hallucinating LLM should be used to choose military targets autonomously.

  10. Assessment: Can’t be checkedSaid

    “Which is to say we shouldn’t be putting our probabilities before our definitions, or even putting specific probabilities on things we can’t define.”

    What we tested: People should not put specific probabilities on AI model consciousness before defining what consciousness means.

Held from this post (5)

These records did not pass every check, so nothing about them is published. Held is not a finding.