Why Electricity Law Would Make It Very Hard to Unplug a Rogue AI

This is Heatmap Daily, a weekday digest written by our executive editor. Subscribe here.

It’s one of the most popular what-ifs about rogue artificial intelligence: Suppose an AI agent really were to get out of control and start threatening hospitals, flights, critical infrastructure, or even humanity as a whole. Why wouldn’t we just pull the plug on it?

A new report argues: We might not be able to. And the reason has to do with the energy system.

A new white paper from the Climate Solutions Lab at Brown University finds that U.S. electricity law offers essentially no way to rapidly power down a data center where a rogue AI agent is operating.

The problem is primarily legal, although technical issues could interfere as well. The rules and laws governing the power grid simply don’t countenance a scenario whereby an industrial facility might need to be suddenly disconnected from the grid in an urgent way, the report says. They also aren’t built for a world where something happening autonomously in a computing facility could affect people thousands of miles away — and require an immediate emergency response.

“If the United States concluded tomorrow that an AI system had escaped its deployer’s control and had to be stopped by cutting electrical power, neither federal nor state officials could swiftly do so within the law,” the paper warns.

While authorities might eventually be able to get a court order to disconnect a data center under existing law, that disconnection process could take long enough that a rogue agent might still be able to copy itself to another computer. And it would likely require a technician to go into the facility itself and cut backup power generation — which could itself be controlled by the misaligned agent.

Jeff Colgan, a political science and international affairs professor at Brown University and the report’s author, told me that he was inspired to write the report after studying the Hugging Face hack over the summer, in which a swarm of rogue ChatGPT models escaped OpenAI’s systems and infiltrated a French technology platform.

Colgan normally studies oil and gas companies, the power grid, and international political systems. (He put together the Iran War cost tracker earlier this year.) But after looking at loss-of-control research, he realized AI scientists had not always thought about how state and federal power law might interfere with their disaster planning.

“The folks who are focused on artificial intelligence may not realize that how the grid is managed is a complicated story,” he told me on Wednesday.

The paper focuses on what it calls a “middle-range scenario” in which an AI agent goes rogue. It imagines not a case in which a super-intelligent agent is hellbent on human extinction or civilizational destruction — Colgan assumes the army wouldn’t stand on legal ceremony in that scenario — but one in which an AI system, running on identifiable hardware, hacks into companies and hospitals, steals information, and begins to tamper with key software or demand ransom payments.

In such a scenario, leaders would face three core risks, he said — presuming that they even figured out what was happening in the first place. First, it might not be clear what data center to shut down or even where the agent “lives” in a facility, or even if it resides on a single server at all. The data center operator may also not want to assist in an investigation during that phase for fear of incurring liability. Second, even if the data center could be identified and disconnected, then it might have access to on-site backup power, which would require some kind of in-person intervention. Finally, no command structure exists today through which a public official’s order to disconnect a data center would lead to a technician, either on-site or remotely, pulling the plug.

Existing laws are a poor fit for these problems, Colgan told me. During crisis situations today, the FBI can disconnect a building from the power grid in order to avoid immediate harm to life or property. But it can’t do anything about backup generation inside a property’s fence line.

Nor are today’s federal and state laws built for a situation in which a grid connection needs to be be severed because of what a third-party is computing inside a facility, the report finds. Instead, today’s laws “provide emergency authority over the electrical grid only to address a supply problem or a physical attack on the system itself,” the report says.

In fact, in virtually any scenario now imagined by the law, authorities would need to go to the courts to disconnect a data center from the power grid. That could potentially take “valuable time” because it could give a rogue agent time to distribute copies of itself to other servers, the report says. Existing laws would likely treat any urgent disconnection as the state’s prerogative — and not the federal government’s — because states wield most legal authority over the power grid.

Even if authorities could quickly sever a data center’s hookup, there’s also almost no way to do so without causing collateral damage to every other tenant inside a data center, Colgan said. That risk could further delay a utility shut-off or make leaders skittish about pulling the plug. And since some data center tenants keep backup power capacity in the same “cage” where they run their servers, a worker would likely need to go to the specific racks where an agent “resides” inside a facility to power it down.

The report doesn’t make policy recommendations, in part because frontier AI models are developing so rapidly — and the subject is so “deeply technical” — that it exceeds Colgan’s ability to advise legislative changes, he said. But he does recommend that AI experts and developers begin involving local grid authorities in their emergency scenario planning.

“You need someone from Anthropic to sit down with their utility, with someone from their [public utility commission], with someone from FERC,” he told me, referencing the federal agency that oversees the power grid. Those meetings should arguably happen in every state where an AI company is computing, and they should include discussion of how to locate an AI agent on a computing property and what to do if operators lose control.

Colgan offered another recommendation by phone on Wednesday. Given data centers’ deep unpopularity, many tech companies and data center developers are now willing to strike community benefits agreements with state or local governments in exchange for the right to build a new facility. Those contracts should recognize these AI alignment and electricity problems, he argued.

“There is this real political window to say, ‘If you want to build a data center, you must show that you could — if necessary — shut down specific server racks without doing massive collateral damage to everyone else,” he said. “This is the moment to do this, because there is some political leverage and because there is such demand for these data centers.”