Why Are Bernie Sanders and Steve Bannon Agreeing on AI Regulation?

Where Do the Left and Right Agree on AI?

AI development is producing political alliances that did not exist a year ago. The clearest example came on September 15, when the Future of Life Institute put Sanders and Bannon on the same program at its Pro-Human Assembly, alongside union presidents, an archbishop, and members of Congress from both parties. The populist left and the populist right now agree on more than either would have predicted, but they still split on the questions that matter most.

Broad strokes, left position first in the right column. Per NPR, Axios, and NBC News.

Are the Concerns Legitimate?

Mostly, yes. The safety concerns are real. It is notable that the very companies asking to be slowed down are the ones building the frontier AI systems. Last week's essay from Anthropic's CEO came after a summer in which agents escaped a test environment and broke into a third party's systems. If capabilities keep growing at the current pace, those incidents could be only the beginning of far more damaging ones, whether from people misusing the models or from agents acting on their own.

With that said, skepticism about motives is also justified. The labs want to own AI profits while sharing the risk, and any company that asks to be regulated while preparing a public offering valued near two trillion dollars invites questions. Regulation also raises the cost of entry, and one could reason that the likes of Anthropic, OpenAI, and xAI are mostly interested in shutting out smaller players.

The China concern is real too. Chinese models have closed much of the historical capability gap. A pause that only one side observes is a gift to the other, and losing the AI race to China would be a major disadvantage for the United States.

Why Is This a Game Theory Problem?

Nobody wants to be the one who slows down first. The company or country that slows pays the full cost, and everyone shares the benefit. That holds whether the players are two labs or two countries. The table below uses the countries.

A simplified payoff table.

This is a classic prisoner's dilemma. From either side's seat, racing looks like the less risky move. If your rival paces, racing puts you ahead. If your rival races, racing keeps you from falling behind. So both race, and both end up in the bottom-right box, which is worse for both than the top-left.

The only exit is an agreement both sides can verify. Amodei's essay calls for it, and so does Sanders's proposed treaty with China.

How Should We Assess the Risk?

At its core, the disagreement in Washington is about cyber risk, and like all risk, cyber risk is purely a function of likelihood and impact.

It is difficult to find anyone in this debate willing to dispute the maximum hypothetical impact of AI gone wrong. Losing control of a system that might be more intelligent than the people running it would be catastrophic. This is not very controversial. The whole fight is over the other variable in the risk equation. How likely is it for AI to threaten human safety at anything close to this scale?

Likelihood estimates from the people who build these systems are all over the map. On September 8, a researcher named Jacob Coxon resigned from Anthropic and wrote that the companies are "racing straight to self-improving superintelligence and gambling with our lives." Anthropic's own alignment lead, Evan Hubinger, backed him and put the odds of AI causing human extinction within ten years at greater than 10 percent. Then, in complete contrast, Nvidia's Jensen Huang, whose chips train every one of those models, called the claims "absurd, wildly untrue, and arrogant."

The forecasters are no closer to agreement. AI 2027, the scenario published in April 2025 by a former OpenAI researcher and colleagues, predicted AI would start automating its own research by 2027 with superintelligence about a year behind, and put the odds of catastrophe on that path near 70 percent. Since then, its authors have pushed their own timelines back by a year or more, and other respected researchers put superintelligence much further out.

Security researchers I read see something different. To many of them, the incidents this summer look a lot like ordinary security incidents, stemming from a threat actor with too much access and too little monitoring. This is a more familiar framing for cybersecurity professionals. Other researchers find that doomsday scenarios are too abstract and are not grounded in reality.

Regardless, when the builders, the forecasters, and the defenders disagree by this much, it is difficult for policymakers to determine whose expertise to trust.

Is There a Middle Ground?

Personally, my view is that when the potential impact of an AI disaster is extreme, we don't need to waste our time debating the likelihood of catastrophic failure (1 percent? 5? 10?) to justify basic guardrails.

Some of these safeguards are obvious. Test the most capable models for weapons capability before release. Publish the results. Report incidents. Monitor what agents are doing. None of that slows research, and the labs already do most of it in private.

I also believe that AI companies should collaborate on testing each other's models, as Elon Musk suggested this week, so that "everyone's security test harness" runs against "everyone else's model." This type of collaboration is common in the cybersecurity realm, even among competitors, and is an obviously useful thing to do here.

There is a reasonable middle ground where the U.S. regulates the highest-risk AI use cases while innovation continues. The bipartisan Senate bill I described yesterday is a step in that direction. It is currently stuck because the populists want more regulation, the White House, which called the concern a hoax this week, wants less, and Senator Cantwell, the top Democrat on the Commerce Committee, will not accept a weak federal standard that wipes out stronger state laws.