Gary Marcus Recommendation or value judgment First said 2023-02-17

“And finally, for the executives and chief scientists of OpenAI, Google, Microsoft, and other companies engaged in building these things: DO provide detailed information about how your systems have been trained and updated., and DO provide a method whereby a user can mark a test example or benchmark so that it is not used in training and thus made useless for testing. To fail to do that is both socially and scientifically irresponsible.”

How Not to Test GPT-3, paragraph 56. The full statement is at the source.

Our reading, the sentence that was checked:

OpenAI, Google, Microsoft and other AI developers should disclose how their systems are trained and updated, and should let users mark test examples to exclude them from training; not doing so is irresponsible.

Assessment: Can’t be checked Not checked against evidence

Where and when Gary Marcus wrote it

Marcus on AI, How Not to Test GPT-3, paragraph 56.

Said: 2023-02-17
Source observed: 2026-10-29live page, full text, public
Evidence: Not checkedboth screens found it cannot be checked
Assessment recorded: 2026-10-29assessment marcus-how-not-to-test-gpt-3-fe4e01-a1

We recorded 12 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 2023-02-17. This page does not claim the wording was captured then.

What the claim commits to

Kind: Recommendation or value judgment
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

Nothing beyond the statement itself.

Our interpretation

None added.

This claim is not graded on evidence. The record checks that the words are quoted correctly and attributed to the right person, and that the claim is classified correctly.

Current assessment

Not graded on evidence

Outcome: Can’t be checked

Both AI providers screened this claim on their own and found that public evidence could not settle it. It is recorded and attributed, not scored.

Why it can’t be checked

Both screens gave the same reason:

A recommendation or a value judgment.

Evidence: Not checked
Deadline state when assessed: No horizon
Assessment recorded: 2026-10-29
Policy version: 2026-10-28.7

Rationale

Not checked against evidence. Two independent screens, one from each provider, both found this a recommendation or a value judgment.

Limitations

None recorded.

Evidence (0 items)

No evidence was searched for. Both screens found that public evidence could not settle this claim.

This claim is about OpenAI, and models from OpenAI worked on this record: gpt-6-astra screened it. Agreement between models is a check, not proof.

This claim was not checked against evidence. claude-opus-5-5 (Anthropic) extracted it from the post and wrote the sentence at the top, our reading of the claim. claude-opus-5-5 (Anthropic) and gpt-6-astra (OpenAI) then screened it independently, without seeing each other’s answers. Both confirmed that the sentence says what the author said, and both found that public evidence could not settle the claim. No model wrote the assessment: code composed its text from the two screens, so there was no separate check of the final text. Code, not a model, decided whether it could be published.

Agreement between models is a check on the record, not proof that the assessment is right. Anthropic and OpenAI make these models, and some claims on this site are about those companies. How the method works.

Where it was said (1)

Said once in the sources this register covers.

2023-02-17: Marcus on AI, How Not to Test GPT-3
And finally, for the executives and chief scientists of OpenAI, Google, Microsoft, and other companies engaged in building these things: DO provide detailed information about how your systems have been trained and updated., and DO provide a method whereby a user can mark a test example or benchmark so that it is not used in training and thus made useless for testing. To fail to do that is both socially and scientifically irresponsible.
Where
paragraph 56
Published
2023-02-17
Source observed
2026-10-29, live page
Access
full text, public

History

Assessment versions

Assessment marcus-how-not-to-test-gpt-3-fe4e01-a1: Current. Outcome: Can’t be checked. Not checked against evidence. Public date not recorded.

Recorded events

2026-10-29
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-how-not-to-test-gpt-3-fe4e01-a1.