Latest / MBA Training AI / NERC risk: when an AI dispatch call is confidently wrong

Transcript

NERC risk: when an AI dispatch call is confidently wrong · MBA Training AI

• 0:00
  Host: Welcome to the Leaders Insights Podcast.
• 0:08
  Host: Today's episode, when the AI is confident and the grid goes dark.
• 0:11
  Host: Here's a sentence that should scare every utility executive listening.
• 0:15
  Host: The AI was completely sure and the lights went out anyway.
• 0:20
  Host: Walk me through why confidence is the dangerous part.
• 0:23
  Expert: Because confidence is exactly what an AI system can't back up when it matters.
• 0:27
  Expert: A large language model, software trained to predict the next plausible word, doesn't know
• 0:33
  Expert: things.
• 0:34
  Expert: It estimates what a reasonable answer looks like.
• 0:36
  Expert: And it delivers that estimate with the same smooth certainty whether it's right or catastrophically
• 0:41
  Expert: wrong.
• 0:42
  Host: But a chatbot's been wrong for years and nobody died.
• 0:45
  Host: What changes in a control room?
• 0:47
  Expert: The blast radius.
• 0:49
  Expert: If a retail chatbot recommends the wrong toaster, a customer returns it.
• 0:53
  Expert: If a dispatch model, the system deciding how much power flows where, makes one confident
• 0:58
  Expert: bad call.
• 0:59
  Expert: You can trip a transmission line that overloads the next one and you've got a cascade.
• 1:05
  Expert: That's the 2003 Northeast blackout.
• 1:08
  Expert: 50 million people dark.
• 1:10
  Expert: From a chain of small errors, nobody caught in time.
• 1:13
  Host: So the failure mode is the same, the stakes are just bigger?
• 1:16
  Expert: The stakes cross a legal line, which is the part execs miss.
• 1:20
  Expert: In North America, you have INERC, the North American Electric Reliability Corporation,
• 1:25
  Expert: the body that writes the mandatory rules for keeping the grid stable, violate their reliability
• 1:31
  Expert: standards and you're looking at fines up to a million dollars per day per violation.
• 1:36
  Expert: A wrong toaster doesn't come with a regulator writing you a seven-figure invoice.
• 1:41
  Host: Let me push on the confident bit.
• 1:43
  Host: Isn't the fix just to make the model more accurate?
• 1:46
  Expert: That's the trap.
• 1:48
  Expert: Money and honesty about uncertainty are two different things.
• 1:51
  Expert: You can build a model that's right 98% of the time and still be dangerous because in the
• 1:56
  Expert: 2% where it's wrong, it gives you no signal it's wrong.
• 2:00
  Expert: It doesn't say I'm guessing here.
• 2:02
  Expert: It says dispatch 400 megawatts in the same tone it always uses.
• 2:07
  Expert: The grid needs a model that flinches.
• 2:09
  Host: Flinches how?
• 2:10
  Host: What does that actually look like?
• 2:12
  Expert: You want the system to output a confidence range, not a single number, and to escalate
• 2:17
  Expert: to a human when that range gets wide.
• 2:19
  Expert: Think of it like a pilot's autopilot.
• 2:22
  Expert: It flies the plane, but it's screaming for the captain the moment conditions drift outside
• 2:26
  Expert: what it was trained on.
• 2:28
  Expert: A grid model with no hand it back to the human trigger is an autopilot with the alarm
• 2:33
  Expert: disconnected.
• 2:34
  Host: Google's been pushing AI for grid optimization.
• 2:37
  Host: What do their numbers say?
• 2:39
  Expert: Google's AI teams have claimed their forecasting models can predict wind farm output 36 hours
• 2:44
  Expert: ahead and boost the value of that energy by around 20%.
• 2:49
  Expert: Real work.
• 2:50
  Expert: But worth noting, Google sells the cloud and AI services that run those models so that
• 2:55
  Expert: 20% is a vendor's own figure, not an independent audit.
• 3:00
  Expert: I'd treat it as a ceiling, not a promise, and I'd want to see it replicated by someone
• 3:05
  Expert: who isn't selling the compute.
• 3:07
  Host: So how does a utility actually deploy this without betting the grid on a black box?
• 3:11
  Expert: Two layers.
• 3:12
  Expert: First, the AI advises.
• 3:14
  Expert: A human dispatches.
• 3:16
  Expert: For now, the model surfaces options and it's uncertainty.
• 3:20
  Expert: The operator pulls the trigger.
• 3:21
  Expert: Second, you hard code physical limits the AI cannot override.
• 3:26
  Expert: If the model says push a line past its thermal rating, the point where the wire literally
• 3:31
  Expert: sags and faults, a separate dumb deterministic system says no.
• 3:36
  Expert: The AI proposes physics vetoes.
• 3:39
  Host: Doesn't that defeat the point?
• 3:41
  Host: If a human checks everything, why buy the AI?
• 3:44
  Expert: Because the AI handles the 10,000 boring decisions a second that a human can't and flags the
• 3:49
  Expert: 12 that need judgment.
• 3:51
  Expert: You're not replacing the operator.
• 3:53
  Expert: You're aiming their attention.
• 3:55
  Expert: The mistake is thinking the goal is fewer humans.
• 3:58
  Expert: The goal is humans looking at the right problem at the right moment.
• 4:02
  Host: Give me the one thing an energy exec should do Monday morning.
• 4:05
  Expert: Ask your AI vendor a single question.
• 4:08
  Expert: When your model is unsure, what does it do?
• 4:11
  Expert: If the answer is it always returns an answer, you don't have a grid tool, you have a very
• 4:16
  Expert: expensive gambler.
• 4:18
  Expert: Make uncertainty a visible output or don't deploy it near the switch gear.
• 4:21
  Host: Sources for today's episode.
• 4:24
  Host: TechCrunch AI, the decoder.
• 4:26
  Host: Ours Technica AI, Google AI, vendor, AI lab.
• 4:30
  Host: That's it.
• 4:31
  Host: The AI decision tools are live at MBA-training.com.