In November 2024, the North American Electric Reliability Corporation released a white paper with a sentence that carried the finality of a bell toll: “the ‘genie’ cannot be put back into the bottle” (p. 1). It was not the first time technology had crept into the control room, but it was the first time the institution tasked with reliability admitted, publicly and without hedging, that artificial intelligence would soon sit alongside human operators. The report, AI and ML in Real-Time System Operations, reads as both a warning and a blueprint. It calls to mind earlier moments when machines arrived in spaces once thought irreducibly human: the cockpit, the nuclear control room, even the trading floor. Each time, the results were double-edged—efficiency gained, safety compromised, and only later, after failure and reform, a more stable equilibrium.
The stakes are high. The grid is no longer just an engineering system; it is the beating heart of the digital economy. Artificial intelligence has not only become a tool of operators—it has become a source of demand unprecedented in scale. Data centers are already consuming gigawatts, with projections to double or triple that appetite by 2030. This duality—that AI is both a consumer of power and a potential custodian of reliability—creates a paradox. The machine arrives as both burden and promise.
NERC’s most important insight is that AI cannot be left to improvisation. It requires a contract. Just as in aviation, where autopilot freed pilots from monotony but left them unprepared for crisis, AI in grid operations risks producing “out-of-the-loop” operators (p. 33). The report insists on clarity: is the machine in active control, serving as co-pilot, or offering decision-support? (pp. 30–33). Without such distinctions, the operator is left in a twilight zone, unsure when to trust, when to override, and when to simply watch.
History instructs us here. Three Mile Island was not caused by a single failure but by a torrent of alarms that overwhelmed human cognition. Likewise, in 2009, Air France Flight 447 descended into the Atlantic not because of mechanical defect but because pilots, long habituated to automation, lost the thread of their own situational awareness. The NERC report does not recount these tragedies, but it echoes their lessons: machines change the role of humans, often in ways unanticipated until the moment of crisis. To avoid repeating history, the contract must be written, rehearsed, and enforced.
This contract is not theoretical. It must be embedded in procedure, procurement, and training. In AIxEnergy, I have described this as a “cognitive contract,” in which roles are not left implicit but spelled out in the same language as switching orders. It is a recognition that reliability is not only electrical but cognitive: a joint system of human and machine, bound together by discipline and trust.
Learning in Public
The report’s survey of 47 entities offers a candid snapshot of an industry in motion. Nearly half are still “learning,” while others have implemented AI for load forecasting, renewable prediction, or cyber defense (p. 41). Some admitted they did not know whether their tools qualified as AI at all (p. 45). It is as if the industry is peering through fog, unsure of the contours of the machine already in its midst.
Policy stances diverge widely. Some organizations have banned public large language models outright, citing cybersecurity and reliability risks (p. 44). Others experiment cautiously, allowing AI to draft outage reports or assist in technical writing. The landscape resembles the early days of cloud computing, when utilities treated it first as an intruder, then as a tolerated adjunct, and finally as indispensable infrastructure. AI is poised on the same trajectory. What is banned today will, once hardened by standards and trust, become tomorrow’s baseline.
Trust is not assumed; it is engineered. NERC insists that operators must be able to interrogate why a model produces its outputs, likening it to reviewing state estimator residuals (p. 42). Confidence intervals, drift detection, and anomaly alerts are not luxuries but reliability controls (pp. 49–50). Without them, AI systems risk becoming black boxes, and black boxes have no place in real-time reliability.
Cybersecurity amplifies the stakes. AI is described in the report as a double-edged blade: a tool for defenders but also a weapon for adversaries. The “Morris II” worm study is cited as evidence of AI’s capacity to propagate attacks (p. 29). NERC recommends extending STRIDE and MITRE ATT&CK into AI contexts and adopting MITRE ATLAS to map adversarial tactics (pp. 26–27). For an industry already steeped in CIP compliance, this mapping is invaluable. It renders AI risk legible, not alien, to existing practices.
Yet the challenge extends further. In my judgment, the supply chain of AI models must itself be treated as critical infrastructure. Training data, model provenance, and update management require audit trails. Without them, adversaries will insert poison into the very systems meant to protect reliability. Just as firmware updates are controlled and logged, so too must model updates be scrutinized.
Training the Operator of Tomorrow
Perhaps the most striking emphasis of the report is on training. Operators must not only learn to use AI but to question it, to spot drift, and to override it when necessary (pp. 26–28). High-fidelity simulators are singled out as indispensable, but with a warning: they must replicate the messy ambiguities of reality, not sanitized case studies (p. 34).
This resonates deeply with my own observations. Today’s simulators prepare operators for deterministic contingencies. They do not prepare them for probabilistic systems that present confidence intervals and shifting baselines. The operator of tomorrow must be trained not only to respond to contingencies but to interrogate the machine itself. Did the model drift? Did it misclassify? Did its confidence collapse? These questions must become as natural as checking frequency or voltage.
In AIxEnergy’s work, we have called for integrating hazard layers, congestion indices, and interconnection delays into training environments. Only by rehearsing against structural stresses—blackouts entangled with misbehaving models, congestion compounded by probabilistic errors—can operators be truly prepared. Reliability has always been about preparing for the worst. AI demands we broaden our imagination of what the worst can be.
The Road Ahead
The NERC report’s strengths are evident. It grounds AI in human factors, offers a shared vocabulary of automation modes, integrates cybersecurity frameworks, and calls for transparency (p. 58). Its silences are also clear. The survey is small. The document avoids prescriptive standards. And most glaringly, it omits discussion of AI’s role as a load driver, despite the obvious relevance of AI-driven data center demand to reliability planning (p. 2). That silence is telling. It reflects a choice to confine scope, but also leaves a gap at the heart of the debate.
From my perspective, these omissions underscore the need for follow-on work. NERC has offered a foundation. The task now is to build. Mode contracts must be codified. Explainability dashboards must be standard. AI supply chains must be secured. Competency programs must be institutionalized. These steps align with the EthosGrid™ governance framework I have developed, which insists that machine behavior must be auditable, transparent, and orchestrated according to principles of integrity.
Released in November 2024, the NERC white paper is a landmark, not because it solves the AI challenge but because it acknowledges it. Its insistence on human engagement, explainability, and cybersecurity is both timely and necessary. Its silences remind us that the hardest questions remain unanswered.
For those of us at the intersection of AI and energy, the lesson is clear. The grid remembers. It remembers 1965 and 2003. It remembers storms and near misses. It remembers the costs of complacency. It will also remember how we chose to integrate its new machine mind. If we design collaboration, we may gain resilience. If we improvise, we risk fragility. The contract is ours to write. The future will hold us accountable.
North American Electric Reliability Corporation. AI and ML in Real-Time System Operations. Atlanta: NERC, November 2024.
