When the AI is confident and the grid goes dark

The concept at the center of this article is AI hallucinationAI hallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → in high-consequence decision environments, and the specific confusion worth clearing up is this: most energy executives understand hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → as a text-quality problem, something that makes AI outputs sound odd or include a wrong citation. That framing is too narrow, and in grid operations it is the wrong frame entirely.

Hallucination in a language model is not a bug that appears occasionally under unusual conditions. It is a structural property of how these models generate outputs. They predict plausible next tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → given a context; they do not verify facts against a ground-truth database before responding. When a GovAI research scholar warned, in the wake of an AI hallucination that nearly triggered a US military operation, that "it's important for service members to understand the uncertainty inherent to LLMs," the warning applies with identical force to control room operators, energy traders, and the AI product teams that sell to them.

Why is AI hallucination riskier on the grid than in the office?

Grid operations run on narrow tolerances. The North American Electric Reliability Corporation's standards, particularly the CIP and FAC families, impose mandatory compliance obligations with penalties that reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → $1 million per violation per day. A single dispatch error on a high-voltage transmission line can initiate a voltage collapse that propagates faster than any human operator can intervene. The August 2003 Northeast blackout, triggered by a software alarm failure in FirstEnergy's control room, cut power to 55 million people; the sequence from first fault to widespread outage took roughly 90 minutes. An AI system that hallucinates a grid state, even briefly, compresses that window.

The economic structure of utilities amplifies this asymmetry. Regulated investor-owned utilities like Duke Energy or Dominion Energy earn their allowed return through rate cases approved by state public utility commissions. A reliability incident doesn't just incur direct NERC fines; it exposes the utility to disallowance proceedings where the commission can refuse to let the company recover costs it incurred. If an AI-assisted outage management system issued a wrong switching order and downstream damage resulted, the question of who bears the cost, shareholders or ratepayers, lands in a rate case with highly public and politically charged consequences.

Energy traders operating in wholesale markets under FERC jurisdiction face a parallel exposure. LMP (locational marginal pricing) signals shift minute-by-minute. An AI tool that confidently misreads a constraint on the PJM or MISO network and executes or recommends a position based on a hallucinated grid topology isn't making a small forecasting error; it is generating a liability that regulators and counterparties will scrutinize under market manipulation statutes.

How a stale training set becomes a wrong switching order

The mechanics are worth stating plainly. Large language models trained on text corpora develop strong statistical associations between concepts. When queried about a specific substation's switching sequence or a regional ISO's congestion management rules, the model generates an answer that is linguistically coherent and often partially correct. The problem is that partial correctness in grid operations is not a degraded form of success. A switching order that is 80% right is operationally wrong.

Consider a concrete scenario. An energy management system vendor integrates an LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition →-based assistant to help operators interpret alarm floods during high-load summer conditions. An operator asks: "What is the correct load shedding sequence for the 345kV Allegheny corridor if the Susquehanna-Roseland line trips?" The model has ingested years of grid documentation, past operating plans, and reliability coordinator bulletins. It produces a confident, detailed answer referencing specific substations and sequence steps. But the operating plan was revised in March 2026 following a new PJM planning study, and the model's training data predates that revision. The answer it gives reflects a topology that no longer exists.

This is the operational analogue of what happened in the military hallucination incident: an AI system generating high-confidence outputs from a stale or misrepresented knowledge base, with no internal mechanism to flag the discrepancy. The model does not know what it does not know.Understanding why these confident wrong answers occur at the architecture level is not optional reading for grid AI teams; it changes how they write specifications and acceptance criteria.

Where RAGRAGA method that lets an AI model answer using your own documents, retrieving relevant passages before generating a response instead of relying only on training data.View full definition → belongs in grid AI and where it does not

The tradeoffs here are real and not symmetrical.

AI tools built on retrieval-augmented generationretrieval-augmented generationA method that lets an AI model answer using your own documents, retrieving relevant passages before generating a response instead of relying only on training data.View full definition → (RAG), where the model queries a live, authoritative data source before responding, substantially reduce hallucination risk for factual queries. Several vendors, including AutoGrid (now part of Schneider Electric's portfolio) and Opus One, have built grid optimization tools that constrain model outputs to real-time SCADA or EMS data. When the model can only assert what the live telemetry confirms, the hallucination risk for state-estimation queries drops sharply. This architecture is appropriate for load forecasting support, anomaly flagging, and operator briefing tools where the model surfaces data it has actually retrieved.

What AI tools built on general-purpose LLMs should not do, without significant architectural guardrailsguardrailsRules and controls that keep an AI system inside safe, legal and on-brand boundaries, blocking outputs and actions that cross the line.View full definition →, is generate switching orders, protective relay settings, or load shedding sequences as direct operator instructions. The distinction matters because several current vendor pitches blur this line. An "AI copilot for grid operations" that summarizes conditions from a live data feed is a different product from one that recommends actions based on a prompt. Procurement teams should require vendors to specify exactly which outputs are generated from retrieved real-time data and which are inferred from trained weights.

The NERC CIP-014 physical security standards and the emerging FAC-002 planning standards do not yet address AI-generated recommendations explicitly, butthe compliance exposure when an AI-assisted decision contributes to a reliability event is already a live question in reliability coordinator circles. Regulators don't need an AI-specific rule to find a violation; they need evidence that an entity failed to maintain reliable operations.

The honest tradeoff is this: AI tools add real value in grid environments for pattern recognition, documentation, and scenario modeling under human supervision. They become dangerous when the interface design, or the vendor's sales narrative, positions the model's output as a decision rather than an input to a decision. Operators under stress during a high-alarm event will follow a confident recommendation from a screen. If that recommendation is a hallucination, the confidence is the problem, not the content.

Grid AI deployment needs human override as a hard architectural requirement, not a policy footnote. The military incident should read as a direct precedent: confident AI output, high-stakes environment, inadequate uncertainty signaling. For utilities, the consequence of getting this wrong shows up in the dark.

The full course on this sector:AI in Energy & Utilities.