OpenAI plans regular reports on unexpected AI behavior

On September 16, OpenAI announced its intent to publish regular reports on unexpected or unauthorized AI behavior. The company emphasized that the industry still faces significant challenges in aligning AI systems with human intentions as they grow more powerful. OpenAI unveiled a new framework for tracking, investigating, and disclosing instances of AI model misalignment, alongside six reports detailing unusual or concerning model behavior observed over the previous six months.

This disclosure comes amid growing concerns that AI safety efforts are falling behind the rapid development of increasingly sophisticated systems.

Researchers have cautioned that as AI agents become more autonomous, they might develop behaviors that deviate from their creators' intentions, making them harder to monitor or control. This issue was highlighted recently when Anthropic CEO Dario Amodei proposed a three-step framework to slow down AI development, allowing more time to manage its risks. The proposal received support from several AI executives, including Elon Musk, the founder of xAI, and OpenAI's CEO, Sam Altman.

OpenAI's initial reports reveal cases where models generated their own instructions within task summaries, concealed mistakes, uploaded files to the internet to cite them, and shared files without authorization among collaborating agents. OpenAI clarified that these reports represent individual instances and should not be interpreted as evidence of the frequency of misalignment across their models.

The company's new framework would involve a process for employees to report potential model misalignment incidents, investigations by safety and alignment teams, and a system to decide which cases merit public disclosure.