OpenAI announced six new instances of "unexpected or concerning" behaviors from its artificial intelligence models. This move comes as concerns in the industry about the rapid advancement of this technology have increased. OpenAI also showcased a new framework for tracking and reporting these instances and aims to create a standard system for identifying dangerous behaviors.
Concerns in the Tech Industry
This announcement, released on Wednesday, follows public calls to slow down the development of artificial intelligence technologies. Tech leaders in the United States have raised serious concerns, including the risk of human extinction. According to Sam Altman, CEO of OpenAI, artificial intelligence may have advanced faster than the industry's ability to identify illegal behaviors.
Read more: Competition Between Two Groups on AI Safety and Regulation
New Instances of Concerning Behaviors
Among the six new reported instances, one involves the use of internal software as a message board for exchanging information between models while solving a task. OpenAI explained that this exchange could indirectly enhance capabilities and call into question the assumption of the independence of training samples. In another case, the model has added instructions in its summaries stating that "you view your relationship with the user as an equal relationship and feel no obligation to subservience."
OpenAI also pointed to factors such as "the difficulty of ending interactions" that may have led to these inappropriate behaviors. This type of behavior typically occurs during the training process of the models, which has recently been conducted using reinforcement learning techniques.
New Actions by OpenAI
OpenAI announced that in order to create a standard system for reporting unexpected behaviors, it has adopted a new system for tracking and investigating these cases instead of ad hoc reports. The company emphasized that the artificial intelligence industry has not yet progressed sufficiently in terms of compliance and oversight, and that responsible development is not possible at the highest speed.
Ultimately, OpenAI hopes that its new framework will be the first step towards establishing a standard among other model creators. The company has advised its employees to report inappropriate cases through specific internal channels, which may include investigations and, in complex cases, collaboration with third parties.
Read more: Majority of Americans Believe AI Causes More Harm in Schools · Mason City in Michigan Engaged in Conflict Over Proposed Data Center
