Exclusive-Behind AI industry's call for independent reviews, a fight over what gets made public

NEW YORK, Oct 13 - Major AI firms Anthropic and OpenAI are fiercely protecting confidential information within their evaluations, fearing that revelations could expose trade secrets, according to two sources familiar with the discussions. Anthropic's CEO, Dario Amodei, has stated that external evaluators should be granted the right to publish key findings without any editorial control.

OpenAI concurred, asserting that assessors should enjoy "editorial independence." The companies are concerned that evaluators working within the companies' walls, known as "embedded evaluation," might inadvertently share proprietary information with competitors. To prevent intellectual-property leaks, measures include overseeing embedded teams and prohibiting evaluators from simultaneously working at multiple labs.

However, what labs consider critical intellectual property may intersect with data relevant for judging a model's risk, leading to disputes and necessitating redactions, according to a separate source with knowledge of past interactions with the labs. Experts emphasize that the independence of evaluators is a significant concern as Anthropic and OpenAI determine who can conduct these reviews.

The Reuters report discloses these negotiations between evaluators and AI labs for the first time, highlighting the struggle between transparency and safeguarding trade secrets amid the potential IPOs for companies valued at trillions of dollars. The outcome of these talks will determine how the world scrutinizes AI technology, which some creators warn may surpass human control.