By Rogé Karma Among the many gifts that Donald Trump bequeathed to Democrats ahead of the midterms was allowing them the exclusive political use of the AI backlash. Now that Democrats have decisively won control of Congress, voters will expect them to follow through on their rhetoric. The problem is that Democrats don’t actually have a plan to regulate AI. In fact, they don’t even agree on what problems need to be solved in the first place.

Until recently, federal AI legislation seemed impossible until 2029 at the earliest. Trump has consistently opposed any form of AI regulation, going so far as to pressure Congress to preempt state laws. Although some Republican lawmakers have seemed more sympathetic to AI regulation, almost none would dare defy their party’s leader. Of the roughly 150 AI-focused bills that have been proposed during Trump’s second term, only one—a bill that criminalized the nonconsensual posting of sexual images, including AI-generated ones—ultimately passed.

Tuesday night’s astonishing results have made AI regulation a much more realistic possibility. Democrats could easily pass an AI bill through the House of Representatives, where they will have a huge majority. And with their surprisingly large Senate majority, they would need only a handful of Republicans to get to a filibuster-proof 60 votes. Republicans could be far more likely to vote with Democrats on AI following an electoral shellacking in which voters made abundantly clear that Trump was on the wrong side of the issue. The threat of a veto remains, but as a lame duck, the president might not care as much as he used to about the risk of slowing down the AI-driven stock-market boom.

So Democrats might be able to get something done about AI. They just don’t seem to know what that something should be. This is true even when it comes to the one AI-related issue that, politically speaking, most obviously needs to be addressed: data centers. Voters are overwhelmingly opposed to the construction of data centers in their communities. Several candidates, including James Talarico in Texas, Abdul El-Sayed in Michigan, Sherrod Brown in Ohio, and Annie Andrews in South Carolina, placed their opposition to data centers at the core of their campaign.

But opinions differ widely about what the precise problem with data centers is and what ought to be done about them. Some politicians focused on rising power bills and proposed making data centers pay for their own electricity generation and grid upgrades. Some seized on the fact that many data-center deals are done in secret, and argued that they should be subject to local approval. And others proposed getting rid of data-center tax breaks. A few candidates, such as Gina Hinojosa, now Texas’s governor-elect, and Andrews, campaigned on a full data-center moratoriums. These proposals varied widely in how long they would last and under what conditions they would be lifted.

Even if Democrats did agree on what to do about data centers, federal legislation wouldn’t necessarily be the right way to achieve their goals. Most Democrats have cast the issue as a state and local one. This might explain why Senator Bernie Sanders’s proposal for a federal data-center ban attracted zero co-sponsors in the Senate.

Beyond data centers, the questions get harder still. For starters, the party doesn’t have one shared view of what aspects of AI need to be regulated. Some are worried about the possibility of mass job losses caused by automation, others about AI deepfakes or children becoming overly attached to chatbot companions. Considerable disagreement exists over how to address each of these concerns.

Take “AI safety”—the set of fears that ranges from rogue AI agents carrying out cyberattacks to the possibility of superintelligent systems wiping out all of humanity. Many Democrats believe that needs to be done to prevent AI from causing problems in the real world. But the ideas for how to do this are all over the place, and each has its own drawbacks.

One school of thought focuses on making sure that AI companies can be held legally responsible if and when their models cause harm. But liability is a fix that occurs only after the damage has already been done. As Charlie Bullock, the senior policy director for the Institute for Law and AI, put it to me, “No one can sue if we’re all dead.”Another idea is to require companies to build a “kill switch” that would shut their models down at the federal government’s command. But knowing whether a model is acting in dangerous ways before the damage occurs is difficult.

Earlier this year, a group of rogue AI agents hacked into the company Hugging Face as well as OpenAI’s own internal servers without anyone at the company, let alone anyone in the federal government, realizing what had happened until well after the attacks had occurred. In other words, by the time someone thinks to reach for the kill switch, it may be too late.

Recently, bipartisan momentum has been building to require AI companies to submit to safety audits, managed by either a new federal agency or approved third-party organizations. This would give the government earlier and deeper insight into the safety of AI models before they’re deployed. But even then, knowing exactly how a model is going to behave in the real world is impossible. As AI models have gotten more and more advanced, they have begun to demonstrate awareness of when they are and aren’t being tested.

This leads to yet another idea: Slow it all down. If AI models are getting too smart to control, the thinking goes, then technology companies should pause their development at least until they know how to properly manage them. This is probably the best way to protect against the possibility of a superintelligent rogue AI. But it raises a set of new, difficult trade-offs. The first is geopolitical: Just because the United States decides to throttle its own industry doesn’t mean other countries, particularly China, will do the same.

In such a situation, the dangers of unchecked AI progress would still exist, and America’s main geopolitical rival would control the world’s most powerful new technology. The second is more straightforwardly political: A forced slowdown could cause a stock-market correction, angering voters. Indeed, the S&P 500 fell 4.3 percent today, in part over fears that Democrats will follow through on promises to slow the data-center build-out.

Going from campaign slogans to sensible legislation will require dealing with the hard trade-offs of governing a new technology. Railing against Big Tech and data centers is easy. Figuring out what to do about them is much harder. But one clear lesson from the midterms is that voters have little patience for politicians who don’t even try to deliver on their core promises. Democrats would be wise not to make the same mistake that just helped get them elected.