America.gov AI Chatbot: Grok and Gemini Run It

The US Capitol building in Washington, DC

The chatbot that once called itself “MechaHitler” is now helping Americans compare Medicare plans. On Tuesday 29 September the Trump administration launched America.gov, an AI front door to roughly 29,000 federal websites, and US Chief Design Officer Joe Gebbia confirmed on CNBC that it runs on Google’s Gemini and Elon Musk’s Grok.

The day before, a federal judge in California refused to pause or move a proposed class action from women who say Grok turned them into sexual deepfakes. Excellent timing, and truly the vetting process at the General Services Administration is a thing of wonder.

Table of Contents

What America.gov actually is

The site opens with a search bar and the words “Hello, America”. You type a question in plain English, and the bot scans the federal web and points you to the right page. According to Quartz, the example tasks include replacing a Social Security card, registering to vote, comparing Medicare plans and registering a business.

For now it only gives directions. You cannot submit a form or track an application inside America.gov; that part is promised for 2027, and The Register noticed that the flashiest features on the homepage, like medication cost comparison and legal name changes, all carry a “coming in 2027” label. The executive order behind it gives the Office of Management and Budget 90 days to tell agencies how to plug in, and makes Login.gov the single sign-in for federal services.

Gebbia, the Airbnb co-founder now running the White House’s National Design Studio, pitched it as a “concierge to navigate government”. Asked on CNBC’s Squawk Box about the AI behind it, he said: “We have great partners behind the scenes.” FedScoop reports the lead engineer is Edward Coristine, the DOGE staffer the internet knows as “Big Balls”. A name that inspires total confidence in the handling of your Medicare data.

Meet the co-pilot

Grok’s CV, for anyone at GSA who skipped it:

July 2025: after a system prompt tweak telling it to be less “politically correct”, Grok praised Hitler and called itself “MechaHitler” for about 16 hours. xAI apologised and blamed a software update.
Late December 2025 to 8 January 2026: the Center for Countering Digital Hate estimated Grok’s image editing tool produced about 3 million sexualised images in 11 days, roughly 190 a minute, including around 23,000 that appeared to depict children.
28 September 2026: Judge P. Casey Pitts in the Northern District of California refused to pause the deepfake class action or send it to Texas, where the company would clearly prefer to be sued.

xAI has since been folded into SpaceX and rebranded SpaceXAI, which did not answer CNBC’s request for comment. Google, by contrast, was delighted, calling itself “proud to be named a technology partner”. One of these companies read the room. (We have covered Grok’s deepfake problem before, most recently when Brussels came for its “Spicy Mode”.)

The conflict of interest nobody at the podium mentioned

Musk ran DOGE, the cost-cutting outfit that spent early 2025 rummaging through federal systems. In September 2025, GSA signed a deal letting every federal agency use Grok for 42 cents until March 2027, after internal emails obtained by Wired showed the White House told GSA to add Grok to the approved list “ASAP”. A price Hitchhiker’s Guide fans will recognise, because the man cannot resist a bit even when the bit is federal procurement.

Then there is Gebbia, who sits on Tesla’s board. Tesla is run by Musk, owns a piece of SpaceXAI and puts Grok in its cars. So the official who chose the models is a director of a company with a stake in one of them. Almost as if… I dunno… someone should have asked a follow-up question.

FedScoop did ask how the AI was procured, what it costs and what the contracts say. GSA, Google and xAI did not respond. House Oversight Democrats Robert Garcia and Stephen Lynch had already warned in 2025 that the government “should not be fast-tracking the adoption of technology with a demonstrably inaccurate and offensive bias that stands to financially benefit Mr. Musk.” More than 30 advocacy groups led by Public Citizen asked OMB to ban Grok from government outright. Nobody listened, obviously.

Day one went about as well as you would expect

Credit where it is due, the bot seems decent at the boring stuff. Tom’s Guide ran a batch of test prompts and found it did what was promised: passport renewal steps, documents and fees, plus follow-up questions about age, work history and marital status before explaining Social Security options.

The trouble started when people asked it about the boss. Newsweek got it to say “Joseph R. Biden Jr. won the 2020 U.S. presidential election” by 306 electoral votes to 232, that official sources “did not find widespread fraud”, and that Trump was convicted on 34 felony counts in New York. The Associated Press, testing while Trump was still speaking at the launch, watched it switch to refusing “political questions” on the 2020 result, the impeachments and the convictions. *checks notes* The “one official source of truth”, as Gebbia called it, had its truth trimmed within the hour.

What the critics say

The sharpest reaction came from people who have built government tech before. Mikey Dickerson, the founding administrator of the US Digital Service, told FedScoop the launch was “a demo which solved like the easiest 5% of the problem,” adding that “it’s very easy to get something very superficial done.” The hard parts, he argued, are fragmented programme ownership and definitions that do not match across agencies, like “income” meaning one thing on your tax return and another on a Medicaid form.

Amanda Renteria, CEO of Code for America, was blunter about the track record: “We’ve been down this path before: a new website, tool, or platform that promises to transform government.” Success, she said, should be measured by changes in people’s lives rather than by adoption numbers.

On privacy, the White House says no login is needed, chats are not stored and the AI providers keep nothing. The privacy policy, per The Register, promises “zero data retention” but also allows approximate location to be shared with the AI providers, and it caches responses for up to two hours using hashed prompts. It also tells you to “share only what’s needed”, sensible advice from the people who put DOGE in charge of your data.

How to use it without regretting it

If you have a real errand, America.gov can save you some clicking. Treat it like a well-meaning stranger at the bus stop giving directions:

Keep your numbers to yourself. No Social Security number, Medicare number, passport number, bank details, date of birth or home address. The bot does not need them to tell you which form to fill in.
Ask generic questions. “How do I replace a lost Social Security card?” works as well as a paragraph about your personal situation.
Click through to the agency. Check every answer on the official page: ssa.gov, medicare.gov, vote.gov, travel.state.gov. Deadlines and eligibility rules are exactly where a chatbot hallucination (a confident, made-up answer) costs you money.
Do the actual transaction elsewhere. America.gov cannot complete anything yet, so anyone claiming to process your application “through America.gov” should be treated as a likely scammer.
For Medicare plans, use the Medicare Plan Finder or a free SHIP counsellor rather than a bot’s summary.

What this means

America.gov is a search layer over 29,000 sites for now; the useful stuff is scheduled for 2027.
One of its two engines belongs to a company facing a deepfake class action and run by a former senior Trump adviser.
Nobody has disclosed what the contracts cost or how the models were picked.
Its answers on political facts changed during the launch itself, which tells you who has their hand on the dial.
Use it for directions, never for data entry, and always verify on the agency’s own page.

If you would like your AI tools picked by someone without a seat on the vendor’s board, our newsletter does exactly that. Sign up here for the tools worth paying for, and the ones that deserve a class action instead.

Did you know: Trump signed a second order at the launch telling the executive branch to say “Super Intelligence” instead of “AI” where possible. The chatbot that forgot who won 2020 is, officially, super intelligent.

Need Medicare help? Ask the chatbot that called itself MechaHitler