DeepMind Researchers Develop Google AI Kill Switch, 'Safely Interruptible' To Resist Human Intervention

The AI kill switch has been developed by researchers at the Google DeepMind to address the concerns of AI being dangerous enough to be capable of learning anything wrong once the real-world inputs are acquired by the machines. With the new technology, a framework through "safely interruptible" AI ensures that the machine will not learn to oppose human attempts to intervene in its learning processes.

AI kill switch framework deals with a wing of machine learning called "reinforcement learning," where an agent learns with reward function. This means the agent examined each potential action based on how well it does for a predetermined objective, and that the closer it becomes, the more reward it will get.

However, one drawback is that human programmers may not always expect each potential way there is to get a given reward. A learning agent may find out some short-cuts that can make the best use of the reward for the machine yet may end up appealing to humans. The human programmers may also be able to change their learning algorithm to explain for this, yet ultimately risk the possibility of terminating the reward function entirely, RT reported.

Another problem being looked at is how to interrupt the robot without the machine learning about such interruption, like if the robot should think the interruption will not occur again. The danger is beyond an ineffective warehouse: if the human intervention cannot make use of  the agent's given reward function, it is likely that the agent has learned to prevent and resist the future interventions.

This problem can be associated with corrigible AI agents that recognize that they are basically flawed and under-development so they think of any human intervention as a neutral thing for any reward function. The paper has proven that two common AI learning frameworks are now interruptible, but also suggests a system in which an agent is programmed to look at human interventions as the effect of its own decision-making processes.

AI kill switch is, therefore, a safely interruptible AI that can be shut down anytime, and if the robot is designed with this, then it is also built not to resist any human attempts to push such the kill switch, according to ENGadget.