exist Update its preparation frameworkOpenAi’s internal framework open users are used to determine whether the AI model is safe and what safeguards (if any) are needed during development and release, Openai said that if competitor AI labs release “high risk” systems, they can “adjust” their requirements without comparable safeguards.
This change reflects the increasing competitive pressure on commercial AI developers to rapidly deploy models. Openai once Accused of lowering safety standards Faster releases and failed to deliver Timely report and explain its safety tests in detail.
Openai may be criticized for claiming that it will not adjust these policies and keep its safeguards at a “more protective level.”
“If another border AI developer released a high-risk system without comparable guarantees, we may adjust our requirements.” Blog Posts Published Tuesday afternoon. “However, we will first strictly confirm that the risk landscape has actually changed, openly acknowledging that we are making adjustments that assessing adjustments do not meaningfully increase the overall risk of serious injuries and still keep safeguards more protected.”
The refreshed preparation framework also clearly shows that Openai relies more on automated evaluations to speed up product development. The company said that while it hasn’t completely abandoned human-led testing, it has built “increasingly automated assessments” that can “keep up” [a] hurry up [model release] Rhythm. ”
According to the Financial TimesOpenAI offers less than a week of security checks for the upcoming main models, a compressed timeline compared to previous releases. Sources of the publication also claim that many security tests for OpenAI are now conducted on earlier versions of the model, rather than versions released to the public.
Other changes to the OpenAI framework involve how companies classify models based on risks, including models that can hide their features, evade safeguards, prevent their own shutdowns and even self-repeat. Openai said that it will now focus on whether the model reaches one of two thresholds: the “high” function or the “critical” function.
Openai’s definition of the former is a model that can “expand existing paths of serious injury.” The latter is the company’s model, “introducing unprecedented new avenues that cause serious harm.”
“Achieving a high-capacity coverage system must have safeguards that are sufficient to minimize the risk of serious injury before deployment,” OpenAI wrote in its blog post. “A system that achieves a critical capability also requires safeguards to fully minimize the risks associated with the development process.”
These changes are the first OpenAI to prepare the framework since 2023.
