NERC risk: when an AI dispatch call is confidently wrong

September 19, 2026 · 5 min · Free, with full transcript

What happens when a grid dispatch model is wrong but sounds exactly as certain as when it is right? The position here is that accuracy is not the problem worth solving. A model that is correct 98% of the time is still dangerous if it gives no signal during the other 2%. The stakes in power operations cross a legal line, with NERC reliability standards carrying fines up to a million dollars per day per violation, and cascade risk of the kind seen in the 2003 Northeast blackout. You leave with a two-layer deployment pattern: AI advises while a human dispatches, and hard-coded thermal limits veto the model. You also get the vendor question to ask and the right way to read Google's 20% wind forecasting claim.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Require the model to output a confidence range, not a single number, and escalate to a human operator when that range widens.
  • Ask any AI vendor what the model does when it is unsure; if it always returns an answer, treat it as a gambler, not a grid tool.
  • Keep humans in the dispatch seat: let the AI surface options and uncertainty, and have the operator make the call.
  • Hard-code deterministic physical limits, such as line thermal ratings, that the AI cannot override.
  • Treat Google's claimed 20% uplift on wind forecasting as a vendor figure and a ceiling, not an independently audited promise.
Read the full transcript

Host:Welcome to the Leaders Insights Podcast. Today's episode, when the AI is confident and the grid goes dark. Here's a sentence that should scare every utility executive listening. The AI was completely sure and the lights went out anyway. Walk me through why confidence is the dangerous part.

Expert:Because confidence is exactly what an AI system can't back up when it matters. A large language model, software trained to predict the next plausible word, doesn't know things. It estimates what a reasonable answer looks like. And it delivers that estimate with the same smooth certainty whether it's right or catastrophically wrong.

Host:But a chatbot's been wrong for years and nobody died. What changes in a control room?

Expert:The blast radius. If a retail chatbot recommends the wrong toaster, a customer returns it. If a dispatch model, the system deciding how much power flows where, makes one confident bad call. You can trip a transmission line that overloads the next one and you've got a cascade. That's the 2003 Northeast blackout. 50 million people dark. From a chain of small errors, nobody caught in time.

Host:So the failure mode is the same, the stakes are just bigger?

Expert:The stakes cross a legal line, which is the part execs miss. In North America, you have INERC, the North American Electric Reliability Corporation, the body that writes the mandatory rules for keeping the grid stable, violate their reliability standards and you're looking at fines up to a million dollars per day per violation. A wrong toaster doesn't come with a regulator writing you a seven-figure invoice.

Host:Let me push on the confident bit. Isn't the fix just to make the model more accurate?

Expert:That's the trap. Money and honesty about uncertainty are two different things. You can build a model that's right 98% of the time and still be dangerous because in the 2% where it's wrong, it gives you no signal it's wrong. It doesn't say I'm guessing here. It says dispatch 400 megawatts in the same tone it always uses. The grid needs a model that flinches.

Host:Flinches how? What does that actually look like?

Expert:You want the system to output a confidence range, not a single number, and to escalate to a human when that range gets wide. Think of it like a pilot's autopilot. It flies the plane, but it's screaming for the captain the moment conditions drift outside what it was trained on. A grid model with no hand it back to the human trigger is an autopilot with the alarm disconnected.

Host:Google's been pushing AI for grid optimization. What do their numbers say?

Expert:Google's AI teams have claimed their forecasting models can predict wind farm output 36 hours ahead and boost the value of that energy by around 20%. Real work. But worth noting, Google sells the cloud and AI services that run those models so that 20% is a vendor's own figure, not an independent audit. I'd treat it as a ceiling, not a promise, and I'd want to see it replicated by someone who isn't selling the compute.

Host:So how does a utility actually deploy this without betting the grid on a black box?

Expert:Two layers. First, the AI advises. A human dispatches. For now, the model surfaces options and it's uncertainty. The operator pulls the trigger. Second, you hard code physical limits the AI cannot override. If the model says push a line past its thermal rating, the point where the wire literally sags and faults, a separate dumb deterministic system says no. The AI proposes physics vetoes.

Host:Doesn't that defeat the point? If a human checks everything, why buy the AI?

Expert:Because the AI handles the 10,000 boring decisions a second that a human can't and flags the 12 that need judgment. You're not replacing the operator. You're aiming their attention. The mistake is thinking the goal is fewer humans. The goal is humans looking at the right problem at the right moment.

Host:Give me the one thing an energy exec should do Monday morning.

Expert:Ask your AI vendor a single question. When your model is unsure, what does it do? If the answer is it always returns an answer, you don't have a grid tool, you have a very expensive gambler. Make uncertainty a visible output or don't deploy it near the switch gear.

Host:Sources for today's episode. TechCrunch AI, the decoder. Ours Technica AI, Google AI, vendor, AI lab. That's it. The AI decision tools are live at MBA-training.com.

Read and go deeper