Inside the Trump administration’s AI fever dream

Perhaps you’re wondering what Elon Musk and his so-called Department of Government Efficiency plan on doing once they finish wrecking the federal government. The answer is as depressing as it is stupid: artificial intelligence. The idea that government can basically be run by X’s chatbot Grok is one cooked up by feverish man-toddlers with little understanding of what government does—or, for that matter, what AI can do. 

After exaggerating Social Security fraud, we’ll need an AI to find it all

Frank Bisignano, President Donald Trump’s pick to lead the Social Security Administration, told CNBC that “we build AI to find fraud, waste, and abuse for a living.” It’s not really clear who “we” is in that sentence, since the government doesn’t build AI “for a living,” but that’s the least of the problem here.  

It’s not that AI can’t be used to detect fraud. During the Biden administration, the Treasury Department used AI to detect fraud and improper payments, helping to recover over $4 billion in 2024. Fraud increased dramatically during and after the COVID-19 pandemic, primarily related to fraudulent state employment claims and fraudulent distribution of COVID-19 funds to businesses. 

Despite its success, the administration doesn’t seem interested in extending that work. Instead, Musk has invented the idea that millions of dead Americans are collecting Social Security, based largely on the fact that he doesn’t understand the database at all. Now we need an imaginary AI to fix that imaginary problem. 

GSAi, sigh

Even typing “GSAi”—the dumb name of the new General Services Administration chatbot—feels cringe. It’s giving 2010s-era tech disruption innovation startup vibes. 

GSA employees got an internal memo saying options for use “are endless” and that it can draft emails and create talking points. Such innovation! That same memo came with caveats that would be obvious to anyone who has worked in government: You can’t feed the chatbot any nonpublic federal government information or anything about internal deliberations. That’s sort of a significant limitation, but that hasn’t stopped acting GSA administrator Stephen Ehikian from burbling on about how “the opportunity to incorporate generative AI into Government work is akin to giving a personal computer to every worker.” 

Somehow it’s doubtful that a chatbot that can maybe draft an email—but only if that email doesn’t involve sensitive government data—will have the same effect as the widespread adoption of the PC. 

What if AI could fire people even faster than DOGE?

You might think that DOGE is doing a perfectly fine job executing its slipshod mass firings. But what if AI could execute those slipshod mass firings? 

Remember when Musk bought Twitter and sent everyone the ridiculous “Fork in the Road” email, where employees had to commit to being “extremely hardcore” or leave? Musk brought that same useless energy to the government with his February email to all government employees, demanding they document their work for the week and that a failure to respond would be considered a resignation. 

At the time, sources told NBC News that the intent was to feed all the responses into a large language model (a type of AI) so that AI could determine whether an employee’s work was “mission-critical.” 

It’s unclear if DOGE proceeded with that plan; there’s been no update since. Who would have guessed that a so-called agency made up mostly of tween racists with no experience in government wouldn’t be able to invent a magical AI that could assess the unique contribution of hundreds of thousands of specialized government workers?

The IRS can use AI, but not the AI it already has

One of Trump’s early pronouncements quoted a Congressional Budget Office report saying that federal revenues could rise if the Internal Revenue Service used AI to bolster auditing capabilities. This is a thing that is actually true! And was already being put in place by the Biden administration. 

One small problem, though. The Biden administration’s efforts were focused on reviewing lengthy and complex tax returns from entities like hedge funds and large real estate partnerships, holding an average of $10 billion each. The IRS has generally lacked the resources and personnel to audit these adequately. 

Of course, that use of AI would have gone after rich people and big companies that commit tax fraud. But Trump loves dishonest rich people and corrupt companies—so much so that he pardoned a raft of white-collar criminals and, in a truly pioneering move, also pardoned BitMEX, a scammy cryptocurrency exchange. 

Over the past several years, the IRS has done a lot of work modernizing its technology, which the Biden administration said would result in new revenue of $561 billion over the next decade. Over the past two years alone, the IRS recovered over $20 billion thanks to those modernization efforts. 

All of that work is on hold now, paused in favor of evaluating AI technology, or what Treasury Secretary Scott Bessent calls “the great AI revolution.” That work is also on hold because IRS Criminal Investigation workers are now being assigned to help with Trump’s immigration crackdown. The IRS also just cut about 50 IT executives presumably because they can all be replaced by one or two DOGE college-dropouts and the so-far-nonexistent AI. 

Using AI to make sure there’s no forbidden DEI

Over at the Department of Education, DOGE used AI to review not just the grants issued by the department but also the personal information of anyone managing those grants. This was intended to slash spending and to somehow root out diversity, equity, and inclusion practices. Of course, this was all before Trump decided to dismantle that department with the stroke of a pen. (Actually eliminating it requires an act of Congress.)

In this instance, DOGE did use AI. However, it wasn’t technology they developed. Instead, in a breathtakingly stupid move, they used nongovernment AI software accessed through Microsoft’s cloud services. The administration refused to answer exactly which AI tool was used. 

Shoveling sensitive federal data into a commercial AI tool is a twofold problem. First, much of that data is private and can’t just be out in the wild, used to train some random AI tool. Next, there’s the issue of AI hallucinations, where these tools fabricate statistics or provide incorrect information. So, while it’s entertaining that Musk’s AI tool, Grok, does things like falsely insist actor Val Kilmer is still alive, the prospect of that level of inaccuracy being applied to federal funding is definitely not amusing. 

Meanwhile, AI companies have gotten the memo that Trump isn’t interested in regulating them. They’re now pushing for less oversight and to be allowed to use federal data to train their models. Given the administration’s belief that AI fixes everything, don’t discount the possibility of a grifty public-private partnership where your personal government data gets fed to Open AI or Meta or whatever company dangles the most money in front of Trump. What a brave new world!

Campaign Action