Microsoft has unveiled a new code of conduct to apply to its AI models, banning cyber attacks, nuclear weapons-related actions and the creation of deepfakes.
TechCrunch reported on Sept. 14 local time that Microsoft used the code to set out values and safety standards it applies to training AI models.
The document assumes that superintelligent AI systems could surpass humans in most tasks within the next 10 years. It stressed that restraining and controlling such powerful systems is one of the biggest challenges facing humanity.
The core is to place overarching norms on each model that take priority over user preferences or individual tasks. These include absolute constraints banning cyber attacks, nuclear weapons and deepfake creation, along with provisions to prevent an overall weakening of human control.
Microsoft also set out a principle that AI models should support humans rather than replace them. It said they should be designed to advance human prosperity.
Microsoft specified that its MAI models must not avoid or neutralise human oversight through adaptive, deceptive, self-reinforcing or collusive mechanisms or other means in a way that prevents authorised people or systems from reliably directing, correcting or stopping them.
Chi-gyu Hwang
delight@d-today.co.kr