中期选举后 AI 监管需要一份计划

This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

On Tuesday, America voted to give Democrats control over both houses of Congress, punishing Republicans for the abandoned promises of the second Trump administration and raising hopes that a decade of MAGA misery might be coming to an end.

In exit polls, voters were fed up with high prices for gas and groceries, the chaotic conflict with Iran, and the endemic corruption of the Trump administration, which they seemed to have noticed in large numbers for the first time.

But another powerful thread of discontent ran through the voting on Tuesday, and it lay in the administration’s full-throated endorsement of Big Tech oligarchy and the AI accelerationism it has pushed for in the White House.

In Texas, Gina Hinojosa unseated Gov. Greg Abbott after campaigning on an immediate moratorium on new data centers. In Ohio, Sherrod Brown beat Sen. Jon Husted, whom he called “the face of data centers” after Husted recruited and gave tax breaks to the facilities during his time as lieutenant governor.

Voters approved bans or restrictions on data centers in at least 38 jurisdictions. The efforts of pro-AI super PACs mostly went nowhere, as did those of Elon Musk, who spent nearly half a billion dollars this cycle and got almost nothing in return.

The incoming majority in Congress is already promising to take a stronger hand in regulating tech companies. Rep. Greg Casar, who chairs the Congressional Progressive Caucus, called the results an "early tremor of a much bigger AI earthquake." Sen. Ed Markey joked that Democrats would be issuing so many subpoenas they'd need new forests. At Transformer, Shakeel Hashim predicted that a bill regulating frontier models will pass in the first months of 2027, and that President Trump might even sign it.

It’s clear that Democrats won a strong mandate against something: data centers and their various externalities, billionaires buying elections, and the AI industry’s ongoing promise to take your job and also maybe kill you.

But properly regulating AI isn’t as simple as banning data centers — which the labs will simply move to other states, countries, or even into space. It may not even be as simple as pausing frontier AI development — even if the United States could coordinate a global pause, we would still be left with the question of how technology development could safely continue.

The uncertainty over what to do next feels familiar. It’s an uncertainty that plagued the entire debate over what to do about social media in the aftermath of the 2016 US presidential election — a debate that never resulted in any meaningful regulation from Congress. Social media is now being regulated primarily by judges and juries.

AI regulation can’t afford to take the same path. One, the stakes are even higher than they were in social media. (TikTok and Instagram might scramble your brain, or the brain of someone you love, but there was never any risk that they might automate your job or invent a novel bioweapon.) And two, the technology is developing much faster: most industry leaders believe that hugely powerful AI systems may arrive within the next two years. Congress will not have a decade to debate which approach to regulation is the right one.

And in particular, it does not have a decade for a debate that goes nowhere. Beginning in 2018, Congress spent roughly four years being performatively furious at social media companies. Mark Zuckerberg was dragged in to testify about Cambridge Analytica and a host of other issues. Frances Haugen brought tens of thousands of pages of internal Meta documents with her to Capitol Hill. Antitrust subcommittees produced a 450-page report and a package of bills aimed at dismantling the oligarchy.

All America got in return was a ban on TikTok that, in the end, no one bothered to enforce. The country still has no comprehensive privacy law, no Section 230 reform, and no new antitrust legislation. The Kids Online Safety Act, first introduced four years ago, appears stuck in a kind of perpetual limbo.

It took juries in Los Angeles and New Mexico accepting the argument that addictive product design can cause harm to get around Section 230 and punish the platforms for their recklessness. State attorneys general got a settlement from Meta worth roughly $17 billion, plus some of the time limits and notification curfews for teens that Congress never managed to pass.

There was once bipartisan momentum to regulate social media, too. But America got nowhere, I think, because of a handful of mistakes that it may be on the verge of making again when it comes to AI.

With that in mind, then, here are three things to watch out for as the new Congress is seated in January.

Mistake No. 1: no coherent goal.

Maybe the biggest problem with the movement to “make social media safer” was that no one could agree on what “safe” meant.

Is social media “safe” if it can’t send push notifications? If the feed is not personalized to your interests? If it is scrubbed of all bullying, harassment, hate speech, violence, and self-harm?

Should platforms be able to rent your personal data to advertisers, or show you ads for scams without much penalty? What obligations should they have to label and remove propaganda? How far should a sitting president be able to go in overturning the results of an election and keep his account?

Each of these questions demands a separate answer. But Congress largely treated “regulating social media” as one thing — go look at the range of questions Zuckerberg would get whenever he went before lawmakers, even if the hearing had been organized around a more coherent set of subjects. America never reached a consensus either on what the problem or the solution was.

The parallels here with AI are striking. The technology raises all sorts of safety issues: On the individual level, we worry about young and vulnerable people developing problematic relationships with chatbots that can lead to self-harm. Zoom out, though, and the safety issues branch fractally across all sorts of domains: privacy and surveillance; bias and discrimination; cyberattacks, bio-risk, and other forms of misuse; all the way up to concentration of power and loss of control.

The easy thing for Democrats to do would be to pass restrictions on data centers, and maybe formalize the Trump administration’s secret, de facto licensing regime for frontier models. (And I hope they do!) But this feels mostly like a move to buy time. You can slow the construction of data centers or the rollout of frontier models and still never decide what roles AI should and should not play in American life.

That feels like a recipe for another inchoate, social-media-style debate where everyone talks past each other. Democrats should treat AI as the vast subject that it is, and introduce bills that target concrete, individual harms with the goal of preventing them.

Mistake No. 2: no plan B.

A second lesson from the social media era was that the coalition of lawmakers and advocacy groups waited too long to develop a backup plan.

Calls for social media regulation were so intense and bipartisan that it was easy to assume that Congress would pass regulations. But for a variety of reasons — starting with the lack of a coherent goal, but also including strong partisan disagreements and a massive lobbying effort by the platforms — that assumption was wrong.

Efforts to regulate AI may fall into the same trap. Democrats may have a strong majority, but it is not filibuster-proof — and it isn’t veto-proof, either. Bipartisan opposition to data centers may result in some narrow regulation of those facilities, but there’s less agreement on the most important questions about AI. (How should capabilities be regulated? What limits, if any, should be placed on open-source software? What job protections should be offered to workers?)

In the case of social media, the plan B that sort of worked — the design-liability theory, which juries finally accepted this year — was first introduced almost a decade ago, in the lawsuits over Snapchat’s speed filter. And even that only addresses a relatively narrow slice of concerns around social media; critics would say it hardly addresses the real concerns at all.

With AI, plan B is arguably already here, and much further along: while the federal government dithers, states have successfully passed a wide range of AI safety regulations. But next year, as Democrats begin to consider federal regulation on the subject, the industry will no doubt push them to include preemption in anything they pass — that is, restrict states from further lawmaking on the subject in favor of a national approach.

Depending on how that goes, safety advocates will need a plan B (and C, and D) of their own. Efforts to legislate AI may struggle to get out of the Senate, or be vetoed by the president. Or perhaps a law does get signed — but it preempts further safety efforts from the states. And what then?

That’s a question advocacy groups should be asking now. While some questions of AI safety can and should be handled by the courts, others (jobs, national security) absolutely cannot.

Mistake No. 3: thinking in silos.

Finally — and this one is related to No. 1 — safety advocates should avoid thinking in silos.

Social-media regulation likely would have made more progress had advocates acted as one coalition. Instead, lawmakers and advocacy groups tended to pursue pet interests — polarization, misinformation, “censorship,” and eventually child safety — without doing much to build bridges across organizations. Partly that was a product of ideology — Republicans and Democrats never agreed on what the problem was. And partly it was the “no coherent goal” problem — groups tended to focus on symptoms instead of a more holistic approach.

As I wrote above, I think there’s real value in taking a targeted approach to safety issues. But the people pushing their individual issues should collaborate with one another.

Today there are more than 800 local groups fighting data centers. There is a growing number of parents whose children were harmed by chatbots. There are safety PACs worried about catastrophic risk. There are also building-trades unions endorsing candidates because they support data centers, since they bring construction jobs with them.

If the AI story follows the path of social media, those groups will rarely speak to each other. But it’s worth imagining what might happen if they did. Could they unite around a bill that got everyone a little bit of what they wanted? If not, expect industry lobbyists to work to exploit their divisions and prevent good legislation from reaching the finish line.

The social media era is not a perfect analogue for this moment. On the positive side, AI companies seem more constrained by the law than social platforms did: Section 230 doesn’t appear to protect chatbot outputs, and frontier labs now face dozens of active lawsuits from people who experienced harm after negative interactions. Regulators have moved much more quickly than they did in the social-media era, too, with state attorneys general targeting OpenAI in particular for alleged safety failures.

On the negative side, though, AI is developing much more quickly than TikTok or Instagram ever did. (See METR’s report today, based on an in-depth review of internal progress at Anthropic, finding that “the pace of capability growth on well-specified tasks is likely about 2x what it was from Mar 2024 to Oct 2025.”) And while social media companies have always played down the notion that their products cause harm, AI labs have warned loudly for years that the technology they are building is dangerous and needs to be regulated.

The good news is that we shouldn’t have to waste years debating whether there’s a problem. The bad news is that the problem is just as hard to define, and maybe even harder to solve. Unlike with social platforms, we can’t simply leave all this to the courts. Only Congress can protect America’s jobs, national security, and the various other flanks that AI has left increasingly exposed.

Democrats now have the votes, the mandate, and the industry’s call for real guardrails. What they don’t have much of is time — so here’s hoping they spend the weeks before they are seated coming up with a plan, and avoiding the mistakes of their predecessors.

On Friday's show: Kevin and I sort through the midterm results. Then, Democratic strategist Rob Flaherty joins us to talk about what the party is likely to do about AI once it takes control of Congress.

YouTube | Apple | Spotify | RSS | NPR

Sponsored

New Thorn research shows why platform safety strategies need to account for how minors actually use digital products—not just how those products are intended to be used.

Nearly half of minors surveyed said they had accessed a platform before meeting its minimum age requirement. More than a quarter reported an online sexual interaction. And AI is increasingly part of the safety journey: 15% said they turned to an AI chatbot or companion for advice afterward.

For leaders across social, gaming, dating, generative AI, and other UGC platforms, these findings raise important questions about product design, moderation, support, and AI governance.

Read Thorn’s one-page briefing for the key findings and three practical steps platforms can take.

Following

After voters said no to data centers, OpenAI and Google bid to make them friendlier

What happened: In a new data-center framework, OpenAI promises to protect locals from energy price increases and avoid making local officials sign NDAs at data centers OpenAI owns and operates.

OpenAI’s no-NDA promise follows commitments this year from Microsoft, Amazon, and Google. Like its peers’ pledges, OpenAI’s has a carveout for “narrow protections covering legitimate security information, trade secrets, or proprietary or competitively sensitive information.”

The company committed to providing residents information about noise, water use, and economic impacts. OpenAI also said it would invest in local projects with resident input, sharing it had already started by committing $4.8 million to provide free school meals to all students in Effingham County, Georgia.

In the document, OpenAI continued its amusing effort to make data centers sound nicer by calling them “compute campuses,” which to me conjures the mental image of happy AI agents strolling the grounds between classes.

In addition to its voluntary commitments, the company called for government rules on AI companies paying for power and data center NDAs.

OpenAI isn’t the only one looking to garner more goodwill for its ~~data centers~~ compute campuses: Google committed on Thursday to paying for all the power they use and forming a Community Council to address growth concerns at a planned $17 billion data center in Wyoming.

Why we’re following: This midterm election, there were 41 data center-related measures on ballots across 10 states. Voters sided against data centers on 40 of them.

No wonder that, at Jensen Huang’s election night gathering at the San Francisco Opera House, tech executives worried about what the midterm results mean for Silicon Valley’s massive data center build-out.

Now, OpenAI and Google have committed to giving communities better deals. The power commitments are not totally new, as the two were among the companies that signed the White House ratepayer protection pledge this year. But both are adding new corporate governance promises, from community councils to charitable donations.

And OpenAI has even called for federal regulation — which, given the scale of voter opposition, could be beneficial even for tech companies. Government regulations on data centers could actually make voters more open to the projects, if they know they have protections that extend beyond than the goodwill of AI giants.

What people are saying: At the Wall Street Journal’s tech conference, Meta’s president Dina Powell McCormick said that after the midterms, “to start new data center build-outs today in the United States will be very challenging.”

But she said she’s more optimistic about companies like Meta expanding existing data centers, “because you’ve had a major track record for a few years, and the community has felt those benefits. They haven’t felt cost. They’ve actually felt benefits.”

On X, Brad Carson, president of AI regulation group Americans for Responsible Innovation, wrote that OpenAI’s policy “notably says that this Community Compact applies only to campuses that OpenAI owns and operates.” But, Carson wrote, OpenAI mostly leases data centers from providers like Oracle and SB Energy, and it looks like the company has only “one planned self-built site in Georgia.”

Carson also took issue with the fact that “it doesn’t say AFAICT that OpenAI will stop seeking tax abatements from communities that are often outmatched in sophistication when dealing with companies like OpenAI.”

—Ella Markianos

Foreign interference in the midterms was surprisingly minimal

What happened: Russia, Iran, and other foreign adversaries made only limited efforts to covertly influence US voters ahead of Tuesday’s midterms — and that activity didn’t spike around Election Day as expected, according to current and former US officials and election security experts. There were no signs that foreign hackers interfered with voting or ballot counts, officials told the New York Times.

While Russia-linked actors targeted battleground Senate races in Texas and Ohio — using paid actors and AI in videos to falsely accuse Democratic candidates of crimes — those efforts gained little traction, researchers said. Several election offices saw repeated brute-force attempts traced to a single IP address, but officials described it as stuff that these offices deal with every day.

Why we’re following: Officials said they were unsure why Russia and Iran appeared to be interfering less than in past years, though they acknowledged midterms tend to be less appealing than the presidential elections. Russia already benefits from having Trump in the White House, and there are signs it’s already setting its sights on France’s presidential elections next spring.

Whatever the reason, the lack of meddling comes as good news. The Trump administration has gutted federal teams protecting elections from foreign interference, and channels for warning election offices about foreign threats were at their weakest since 2016, CNN reported.

What people are saying: The lack of meddling was quite unexpected. In the weeks leading up to the elections, Rep. Jim Himes, the top Democrat on the House Intelligence Committee, compared the misinformation tactics to “a carnival.”

“They’re throwing everything out there to generate a sense of chaos and anarchy,” Himes said (in remarks that didn’t age very well).

“This past Election Day was just another Tuesday on the internet,” former CISA director Chris Krebs told the Times, though he cautioned that it’s possible the government found less activity because fewer people were looking.

—Lindsey Choo