Anthropic, OpenAI sought control over what AI evaluators can publish, Reuters finds
Leaked details of closed-door talks show the two labs pushing for editorial control over outside safety reviewers even as they publicly backed independent evaluation.
What to know
- Reuters reports, for the first time, details of private negotiations between AI labs and the independent evaluators meant to scrutinize their models.
- Anthropic and OpenAI publicly called for evaluator editorial independence but privately sought significant control over what embedded evaluators could publish, per sources.
- Proposed anti-leak measures include lab staff observing embedded evaluation teams and barring evaluators from working simultaneously at rival labs.
- The outcome will shape how a technology its own creators warn may be advancing beyond human control gets independently scrutinized, amid labs' pursuit of trillion-dollar valuations.
“editorial independence”
OpenAI, AI lab · Channel News Asia / Reuters ↗ · Oct 12
Dario Amodei Anthropic CEOAnthropic AI lab, maker of ClaudeOpenAI AI lab, maker of ChatGPT
How it unfolded 1 development · click the chart to see its coverage articlesposts
-
1
Anthropic and OpenAI respond to Reuters' findings
Anthropic said the characterization did not reflect all its discussions, saying it had made clear that evaluators control their methods, criteria, conclusions and what to publish, while reserving approval rights for new evaluator firms. OpenAI said it was not seeking approval rights over full report text, citing a standard practice of redacting confidential information.
“without editorial control…”
— Dario Amodei, Anthropic CEO · source -
background
Reuters reveals the dispute over evaluator publication rights — Reuters published an exclusive detailing previously unreported negotiations between the AI labs and independent evaluators, describing proposed measures like having lab employees observe embedded evaluators and barring evaluators from working with competing labs simultaneously.
-
background
Anthropic's talks with evaluators begin to shift — By October 8, talks between Anthropic and some evaluators appeared to have evolved from the earlier push for strict control, according to one person familiar with the matter.
-
background
Labs sought strict editorial control over evaluator reports — As recently as October 5, Anthropic and OpenAI were seeking strict editorial control over what embedded evaluators could publish, according to two people familiar with the discussions.
-
background
Amodei and OpenAI publicly back evaluator independence — Anthropic CEO Dario Amodei said external evaluators should be able to publish key findings "without editorial control," and OpenAI agreed days later that assessors should have "editorial independence."
Also covered reported alongside — the timeline has no entry for these yet
-
2 outlets Exclusive-Behind AI industry’s call for independent reviews, a fight over what gets made public
first by Investing.com News, 11h ago · also Channel News Asia
and 1 smaller piece