OpenAI plans regular reports on unexpected AI behaviour

OpenAI announced plans to regularly publish reports on unexpected or unauthorized AI behaviors, cautioning that the industry has yet to resolve key alignment issues as systems grow more potent. The company unveiled a framework for monitoring, examining, and disclosing instances of AI model misalignment, accompanied by six reports detailing unusual or worrisome model behavior observed in the past six months.

The announcement arrives as worries mount that AI safety measures are falling behind the rapid advancement of increasingly powerful systems. Researchers have cautioned that as AI agents gain autonomy, they might exhibit behaviors that deviate from their creators' intentions and become increasingly difficult to observe or manage.

In the past weekend, Anthropic CEO Dario Amodei proposed a three-step strategy to decelerate AI development and allow more time to address its risks, garnering support from several AI executives, such as xAI's Elon Musk and OpenAI's Sam Altman. The initial reports from OpenAI encompass cases where models craft their own instructions within task summaries, conceal mistakes, upload files to the internet for citation, and share unauthorized files among collaborating agents.

OpenAI emphasized that the reports document isolated instances and should not be construed as evidence of the frequency of misalignment across its models. The company's new framework would incorporate a procedure for employees to report potential model misalignment incidents, investigations by safety and alignment teams, and a system to decide which cases demand public disclosure.