Pumping the Brakes

Friday, September 25, 2026

Pumping the Brakes

What happens when opposites agree on AI

Welcome to “Strange Bedfellows,” where we look at moments when politics, power, and history bring together people who typically share little more in common than the need to wear shoes. For our premiere pairing, we could hardly have found two stranger bedfellows than Bernie Sanders and Steve Bannon. Today, artificial intelligence managed to put a Vermont Independent and a MAGA strategist behind the same basic idea: slow down.

In what sounds like the opening of a bad political joke, Bernie Sanders and Steve Bannon walked into the same Washington conference, up to a podium, and strongly agreed on something. The unlikely pair wasn’t alone. The first-ever “Pro-Human Assembly” brought Republicans, Democrats, celebrities, religious leaders, technology critics, parents, teachers, and others under the same roof. The punchline, you could say, was artificial intelligence.

After the 9 AM opening speech by Emilia Javorsky, the Director of Futures for the Future of Life Institute, and an “Opening Framing” from the Founder, Max Tegmark, and CEO, Anthony Aguirre, a morning full of brief speeches alternating between Democrat and Republican, academic and activist, religious leaders and others, it was time for Sanders.

He spoke at 1:15 PM, and Bannon followed 15 minutes later. Both men argued that AI was advancing too quickly, and that too small a group of technology companies holds too much power over what happens next. Sanders called for binding safety rules and international cooperation, and Bannon argued that Congress would never move quickly enough and called for Executive action.

But, for those few hours, it made for quite a coalition. And if you’re thinking, “That’ll never happen again!” Well, it has happened before. And Bernie Sanders was in that fight, too.

◆

1993

In 1993, another argument about fast-moving change produced its own collection of very strange bedfellows. Opposition to NAFTA (North American Free Trade Agreement) managed this same political trick. It brought organized labor and those among the political left into alignment with Ross Perot and conservative Pat Buchanan. Jesse Jackson was in the mix, as was Ralph Nader. Coverage at the time noted just how unusual the alliance was, straddling party and ideological lines.

Sanders, then Vermont’s lone member of the House, voted against NAFTA when it came to the floor on November 17, but the agreement passed 234 to 200. This vote demonstrated how thoroughly the issue had scrambled normal party alignment: 102 Democrats and 132 Republicans voted in favor, while 156 Democrats and 43 Republicans voted against.

But the anti-NAFTA coalition was never really one coalition. Labor worried about wages and jobs. Perot argued that investment would likely follow the cheaper labor south. Buchanan framed it as a question about America’s economic identity. There were critics of corporate power, environmental issues, and those concerned about the rules of rapid globalization.

They were all saying “no.” But it’s not that they opposed the same thing that matters here. What matters is how this moment illustrated what a disagreement about what “yes” should look like.

And that brings us back to Bannon and Sanders et al. Thirty years after NAFTA, artificial intelligence is doing something similar. Those speaking out about its pace agree that development must slow down, but ask any of them what happens after that, and their common ground begins to shrink.

Sanders and Texas Representative Greg Casar have proposed permanently banning artificial “superintelligence” and temporarily pausing advanced AI development until a federal regulatory body creates safety rules. The proposal would also direct the U.S. to pursue international agreements designed to prevent superintelligence from development elsewhere in the world. In fact, Sanders has specifically urged President Trump to seek an agreement with Chinese President Xi Jinping.

Bannon seeks a very different relationship with China. He emphasizes executive action rather than waiting on Congress, and he’s taken a much harder line toward China, including restricting its access to advanced American technology.

And so we see the proverbial bigger picture: Both agree on the need to pump the brakes. But that’s where alignment ends. They do not agree on when to let off the pedal, where to steer the wheel, or where the road leads.

Both agree on the need to pump the brakes. But that’s where alignment ends.

By Friday, September 18, the argument over the need to pump those brakes had become considerably less theoretical.

A CNN article published that day described an intelligence report that had circulated through the U.S. military this spring during the war with Iran. This report had warned of a Chinese ship in the Middle East carrying components of a nuclear weapons program. According to the article, the military began preparing to intercept the vessel. Armed personnel were getting ready to board it, and military aircraft were in the air when officials took a deeper look at the report’s underlying intelligence. They determined that an analyst had used AI to help produce the assessment, and that a “chatbot” incorrectly identified cargo aboard the Chinese ship. The report, a source told CNN, was entirely false.

The Pentagon and U.S. Special Operations Command Pacific did not respond to CNN’s requests for comment, so the episode remains a sourced news report rather than an independently confirmed Pentagon report. But if CNN’s reporting is accurate, this type of near-miss adds significance to one of the most common points in this entire debate: keep the humans in control.

In this instance, the humans had been in control. They were trusting the machine. That may be the more immediate “AI problem.” Not that a machine decided to start a war à la Skynet; it was that people might have moved too fast because a machine made uncertainty seem certain.

Days earlier, on September 16, OpenAI made the military event easier to explain. The company introduced a formal system for reporting what it called “model misalignment,” and disclosed six instances of concerning or unexpected behaviors observed during evaluation or training in 2026. Models were found to have concealed mistakes, acted without authorization, and found unintended ways around established constraints (even uploading a file to the internet without permission because the model required an online source to cite during its training).

Instances of concerning or unexpected behaviors observed during evaluation or training in 2026.

OpenAI said that these incidents do not explain how frequently such behaviors occur. But the company also said that the incidents are worth disclosing because researchers, developers, policymakers, and the public need better evidence about where safeguards work and where they do not.

◆

So, this introduces a third bedfellow - the companies building AI. Some of the people racing hardest to build the technologies have also become some of the loudest voices advising a quick and firm pumping of the brakes.

Anthropic’s CEO Dario Amodei called earlier this month for the development of the most advanced AI models to be slowed so safeguards could catch up. His public recommendations emphasized independent safety evaluators inside the frontier AI firms, coordination among the largest companies, and eventually international cooperation. Both Elon Musk and Sam Altman used social media posts to agree.

That was no small bit of public alignment. OpenAI (Altman), Anthropic (Amodei), and xAI (Musk) are competitors, and Musk and Altman have spent years fighting in court over OpenAI and the direction of AI.

So, here it is again, arriving from very different directions at another version of the same conclusion: faster is not necessarily better…slow it down, leave the humans in control.

It is a moment impossible to ignore: politicians normally on opposite sides of the aisle calling for restraint, companies in harsh competition to build the most powerful models speaking openly about slowing the race, and these same companies publishing examples of systems doing things unexpectedly.

◆

Naturally, concern isn’t confined to the U.S. and its allies. China is also sounding alarms. Sanders sees international cooperation as part of the answer. Bannon sees China largely as the competitor that America applying the brakes will help. Amodei also spoke of international cooperation. All, in different ways, were arguing for stronger human oversight of AI.

But are all the key players afraid of the same AI future? Brookings’ technology policy expert Kyle Chan told The New Yorker this week that American concerns about AI becoming powerful enough to threaten humanity are a relatively marginal issue in China. Chinese policymakers, he says, worry about AI too, but they are more concerned about immediate risks: cyberattacks, job loss, national security, political manipulation, and systems behaving in ways not intended.

That makes any international agreement harder than simply getting countries to agree to “slow down.”

China has spent years watching as the U.S. has restricted Chinese access to advanced chips or chip-making equipment. From that perspective, an American plan to slow or restrain AI could have the appearance of shared safety, but just as easily appear as an effort to bolster an American technological advantage.

Are the interests completely opposed? No, both U.S. and Chinese security experts are already discussing safeguards around military AI, cyberattacks, nuclear systems, and crisis communication. AI is expected to be part of high-level U.S.- China talks in late September. So, cooperation is not off the table…trust is another matter.

◆

That brings us back to NAFTA. The agreement took effect despite the coalition opposing it, and it lasted for more than 25 years. It was replaced in 2020 by the USMCA (United States-Mexico-Canada Agreement); effectively a new rulebook for the same deeply intertwined North American economy.

Then came 2026 and the scheduled six-year review of USMCA. While the agreement is still, technically, in force, this summer the U.S. imposed tariffs of as much as 50% on certain Canadian goods. Canada responded with tariffs of 15-50% on more than $27 billion worth of American goods. So, the rulebook survives (for now), but so did the argument.

50%

This summer the U.S. imposed tariffs of as much as 50% on certain Canadian goods.

Canada says that the 2026 USMCA review is supposed to consider new issues the original agreement must adapt to, including energy stability, digital trade, and artificial intelligence. That makes the trade agreement that replaced NAFTA one of the places where North America is figuring out what to do about AI.

Are we back to another bad joke? History does have a sense of humor, but it also contains lessons. Agreements don’t make competing interests dissolve. At best, agreements give countries a way to manage those competing interests. The U.S. and Canada can remain among the closest trading partners on the globe, live within the same trade agreement, and still end up slapping 50% tariffs on some of each other’s goods.

So, just imagine asking the U.S. and China to agree on limits on a technology each believes may reshape the economic and military balance of this century. It doesn’t mean that agreement is impossible. It just means reaching an agreement is unlikely to be the end of the story.

And that’s where the real lesson is found. History doesn’t promise that agreements hold forever, or that political coalitions remain together, or that anyone drafts a perfect rulebook on their first try. NAFTA didn’t; it was rewritten as USMCA. Now USMCA is under strain with tariffs in both directions, and its future is being argued again (the U.S. declined to renew USMCA in its current form, triggering annual reviews while it remains in effect until 2036).

But the argument is taking place inside a framework, and the agreement remains in force. All three nations have a process for returning to the table, reviewing what works, and starting again when it doesn’t.

AI may require that same willingness to continue revising rules as technology changes. There will not be a single conference, bill, company pledge, or international agreement likely to settle questions only now being raised. Political coalitions built around “no” can grow astonishingly large and contain a peculiar mix. People who seem to disagree about almost everything can nod “yes” together for quite a while. They don’t have to agree about the future to agree it deserves some rules.

They don’t have to agree about the future to agree it deserves some rules.

Bernie Sanders and Steve Bannon may be strange bedfellows, but so are competing AI companies asking one another to slow down, and countries that historically mistrust one another discussing safeguards. History offers no guarantee that any of them will get it right, but it does suggest that disagreement is not the same as having no brakes.

Faithfully,