Independent AI evaluation needs a redaction protocol before the test
Reuters reports negotiations over what outside evaluators may publish. Independence is operational only when confidentiality, redaction and disclosure rights are fixed before results exist.
What happened
Reuters reported that Anthropic and OpenAI negotiated publication and redaction rights with external evaluators, while both companies disputed parts of the characterization or narrowed what control they sought.
Why it matters
An evaluator can have model access yet still be unable to communicate a material failure. Procurement needs a precommitted protocol that separates genuine secrets from adverse findings.
Reuters reported on 13 October that negotiations between frontier-model companies and outside evaluators included control over confidential material, redactions and publication. Two people familiar with discussions said Anthropic and OpenAI had sought strict editorial control as recently as 5 October.
The companies disputed or narrowed that account. Anthropic said the description did not accurately reflect the totality of its discussions and stated that current embedded reviewers control their methods, conclusions and publication, while new evaluator arrangements can include approval rights. OpenAI said it was not seeking approval over full report text, but required protection of confidential information and disclosure of major redactions. Reuters also described prior redactions to an external report about a Hugging Face incident.
This evidence does not show that every redaction was improper or that confidential model details should be public. It shows that “independent evaluation” is incomplete as a label. Independence depends on who chooses methods, sees raw results, decides what is confidential, orders redactions, publishes disagreements and can report urgent risks to authorities.
Sign the publication protocol first
Before testing starts, define narrow confidentiality categories, a redaction log, fixed response deadlines and an escalation route to an independent arbiter. Preserve the evaluator’s right to publish methods, bounded conclusions and the existence and scale of company-requested redactions. Predeclare exceptions for exploitable details and personal data, with a sunset or reassessment date.
Procurement should reject an “independent” badge without the contract. A practical minimum is: evaluator-owned conclusions, visible redaction provenance, protected regulator reporting and a public explanation when the company and evaluator disagree. Link the record to AI auditability; model access alone is not independence.