If there are two things Donald Trump loves more than anything in life, it’s participating in elaborate signing ceremonies and artificial intelligence—sorry, superintelligence. And on September 29, the U.S. president combined both loves with the launch of a new executive order, titled “Streamlining Access to Government Services Through America.gov,” which sets up a new front door through which confused citizens can try to navigate the complexities of the federal government. The America.gov chatbot tries to cut through the thicket of complications inherent in government while catering to the American obsession with AI chatbots. (Half of Americans say they use an AI chatbot, according to the Pew Research Center.)
“We’re simplifying it. We’re glamorizing it,” Trump said at the signing ceremony for the effort to make government work better for people through AI. “We’re making it what it should be.”
And at first glance, the interface—which uses Gemini and Grok as its underlying models and collects a surprisingly small amount of user data—seems pretty good, according to digital government and AI experts.
When Calli Schroeder, senior counsel and director of the AI and Human Rights Program at the Electronic Privacy Information Center, tried out the U.S. government’s new AI chatbot, she did what any good privacy expert would: She tried to break it.
Schroeder asked America.gov about vaccines, immigration, and voter registration—all hot-button issues in the Trump administration. She also pretended she needed a replacement Social Security card for her child and tried to include a made-up Social Security number in the prompt.
But the chatbot, which uses the open-source classifier system Rampart to try to strip personally identifiable information, wouldn’t let her. “That’s great,” Schroeder says. “That’s a really good built-in protection.”
The guardrails weren’t perfect. Other made-up personal information, including a name and address, went through despite the system’s attempt to strip such information from queries or attachments. But America.gov has been described by its creators as “private by design”: It requires no account, does not store chats, and repeatedly refused when Schroeder tried to coax it into giving legal advice.
That matters because users can provide sensitive information even when explicitly told not to do so. When Merici Vinton, a senior fellow at the Center for American Progress who previously helped build the IRS’s Direct File tax service, worked on that program, its human chat service explicitly told users not to provide names or Social Security numbers. People did anyway.
“There should be no savviness” required to interact safely with government, she says.
The chatbot’s usefulness may be more straightforward. The service is free, requires no account, and uses AI to search official government sources and return answers with citations.
“I was expecting to be critical, but it sounds like a very nice tool thoughtfully executed,” says Noah Giansiracusa, associate professor of mathematical sciences at Bentley University and a faculty associate at Harvard University.
Restricting the chatbot to government information gives it an advantage over asking ChatGPT or Claude a question about taxes or Medicare, Giansiracusa says.
“Basically this use of a chatbot is like a search function on steroids,” he says. Instead of simply matching keywords, it can interpret a question and pull together an answer from relevant government websites.
For now, though, that is largely all it does. Queries about whether it’s a crime against humanity to drive a Cybertruck clarify the legal status under U.S. and international law, while other responses avoid overcommitting to claims—important if the chatbot is to become a trusted source of information. That limitation is significant because America.gov adds a new interaction layer on top of existing government data rather than changing how citizens actually complete tasks or interact with agencies.
“It is a cool new chatbot,” Vinton says. “But fundamentally, I don’t get the sense that any of the underlying services or access to those services or access to that data has changed.” It doesn’t change how you file taxes or apply for federal student aid, for instance. “Really it’s just linking to them. It’s something, but it’s not a radical reinvention.”
Dana Chisnell, a longtime government service design expert and former executive director for customer experience at the Department of Homeland Security, makes a similar distinction. She calls the chatbot “a glorified search engine” right now—“which is not a bad thing.”
For someone who has just survived a disaster, for example, being able to describe what happened and immediately find relevant federal help could be genuinely useful. But, Chisnell points out, “making a fresh front door does not change the underlying challenges with helping average residents get services they are eligible for.” Nor does relying on government sources eliminate the question of whether America.gov’s answers can be trusted. Official websites can themselves be incomplete or out of date. As Schroeder puts it, “It’s only as good as the information that’s on the government websites.”
Within hours of the launch, the Associated Press found that America.gov initially answered several politically sensitive questions—including about the 2020 election and climate change—before subsequently refusing to answer some of the same questions.
More fundamentally, some experts question whether an AI chatbot is the best way to make government less impenetrable in the first place. “I’ve never met a user who says, ‘I want a government chatbot,’” Vinton says. “What people say is ‘I want this application to be simpler. I want this response from the agency to be shorter.’” And chatbots, she says, “don’t answer those problems.”