AI Is Moving Into Power Grids and Factories — Scientists Urge Caution

Language models — the AI behind tools like ChatGPT — are leaving the chat window and moving into physical machinery. Power grids, factory robots, self-driving cars, and drone fleets are increasingly using them for high-level judgment calls: decide the goal, pick a route, adjust a target. This new paper surveys that trend across three big areas of engineering and asks a simple question: is this safe?

The problem is that these AI models are slow and a little unpredictable. They take seconds to answer, they sometimes confidently make things up, and they can suddenly go offline. For a power grid or a robot arm, a two-second delay is an eternity. The authors translate AI quirks into classic engineering terms: a slow answer is like a delayed signal, an outage is like a dropped phone call, and a made-up fact is like an unexpected shove that the system must absorb.

Their proposed fix is a two-speed design. The AI acts as a slow supervisor — think of a manager setting the destination and the rules. Beneath it, a fast, mathematically proven controller handles the moment-to-moment work, like cruise control keeping the car steady. The AI never touches the reflexes. That way, even if the AI is late, wrong, or unavailable, the physical system stays stable and safe.

The catch is uncomfortable. The researchers found that as AI models become more capable, formal safety guarantees tend to shrink — the industry is gaining cleverness while losing proof. Almost nobody has mathematically shown these hybrid systems stay stable under real-world conditions. That gap, they say, is the field's biggest unsolved problem, and it matters to anyone relying on electricity, transport, or automated factories.