Researchers Built an AI Grid Assistant That Isn't Allowed to Go Rogue

The people who run electricity grids are having a harder time. As more power comes from wind and solar, the system gets twitchier, and mistakes are expensive. So researchers built an AI helper — the same kind of technology as ChatGPT — that sits in a control room and answers operators' questions. It works on a "digital twin": a software copy of the Greek transmission network that behaves like the real thing. (A digital twin is a simulation you can safely test ideas on.)

The clever part isn't the AI. It's the babysitter. The AI can only choose from an approved menu of analysis tools, can't take more than a set number of steps, and cannot do anything with real consequences without an operator explicitly approving it. Every figure it gives must come from the backend system, with its unit and time attached — nothing invented. That matters because AI chatbots sometimes make things up, and a made-up number in a control room isn't embarrassing, it's dangerous.

The results are striking. On 590 runs, the AI picked the right tool 96.5% of the time and completed the task 93.7% of the time, with all four safety rules holding every single run. A wider test across four different AI models — 1,416 runs — held too, with a statistical floor of about 99.8%. Then the researchers removed the safety layer. The same AI immediately carried out all 45 actions that should have needed approval, and only 39.2% of its answers still had numbers backed by real data.

The guardrails cost 12 to 16 milliseconds per request — a fraction of a blink, invisible to a human. That's the headline: safety here isn't slow or expensive. This is a research testbed on a simulated grid, not a live one, so don't expect it in your local utility tomorrow. But the recipe — let AI suggest, never let it decide alone — is one banks, hospitals and airlines are already copying.