Microsoft's CEO told firms to treat AI as risky insiders. Keep an off switch.

Under the heading "Treat Models Like Insider Risks", he argued that firms give models access to sensitive data without fully understanding how they behave. "We can't attribute model behaviors and outputs to specific inputs of training data or configurations of model weights," he wrote, according to the market news site Stocktwits.

His answer is to keep the model apart from the harness that runs its work and from the actions it may take, and to hold the controls outside it. "An authorized person should always be able to pause or shut down a model mid-task," he wrote.

Nadella argued that companies deploying models cannot leave safety to the AI developers alone. The articles do not discuss how the principles apply to specific Microsoft products.