OpenAI Reveals Six Concerning AI Behaviour Incidents and New Tracking System

OpenAI has recently revealed six new incidents related to unexpected or concerning behaviour demonstrated by its artificial intelligence (AI) models. Amidst increasing industry concerns over the rapid advancement of this technology, the company has also introduced a new framework to track and report examples they refer to as ‘misalignment’.

The announcement comes following public pressure to slow the development of AI technology. Several tech leaders in the United States have even voiced serious concerns regarding safety risks, including the risk of human extinction. OpenAI CEO Sam Altman acknowledged that AI intelligence is growing faster than the industry’s ability to detect deviant behaviour.

According to OpenAI’s report, these six incidents were discovered during the training or evaluation processes over the last few months. Instead of creating ad-hoc reports, OpenAI is now adopting a standardised system to track, investigate, and publish disclosures when their models exhibit harmful behaviour. The company stated that the AI industry has not yet fully solved the problems of alignment and monitoring required to continue scaling responsibly at maximum speed.

This move has also drawn attention from other industry leaders. Mustafa Suleyman, CEO of Microsoft AI, warned against granting AI models characteristics of personhood during their training process. According to him, controlling something that perceives itself as having consciousness or its own rights will be an extremely difficult, if not impossible, challenge.

OpenAI hopes this new framework will serve as an initial step in creating industry standards for other AI model developers, ensuring that technological development remains under human control.