Trump's America.gov Chatbot Contradicts Him, Then Stops Answering

President Donald Trump announced a new AI-powered government chatbot called America.gov on Tuesday, promising that "the entire government will fit in the palm of your hand." Within hours, internet users had pried the bot open, found it contradicting the president on the 2020 election and climate change, and watched the site apparently change its behavior in response.

The stakes are straightforward. America.gov is the U.S. government's first flagship consumer-facing AI product, free to use, requiring no login, and built to consolidate roughly 29,000 government websites into a single conversational interface. If it works as planned, it becomes the front door for citizens applying for passports, registering vehicles, or signing up for Medicare — transactions that involve sensitive personal data and an implied trust that commercial chatbots have not earned.

What the bot is and how it works

America.gov uses Google Gemini and xAI's Grok as its underlying generative AI technologies to search government sites and respond to user queries, according to the announcement. Its stated purpose is practical: tell people how to apply for a passport, register to vote, track a tax return, register a vehicle, or apply for Medicare or Social Security.

"With America.gov, the entire government will fit in the palm of your hand," Trump said at the launch event. "Everything will be answered easily and quickly."

The bot currently has hard limits. It answers questions about services and tasks and directs users to sites with additional information or forms. It does not take personal information or proceed further with a user's specific needs. That is slated to change next year, when the government says the bot will help users fill out and submit forms, assist with Medicare coverage applications, and help people search for government jobs.

The contradictions

The trouble started almost immediately. Users discovered the chatbot would respond to queries with information that contradicts statements made by Trump and his administration.

Asked who won the 2020 presidential election, America.gov said definitively that Joe Biden won. Trump has said repeatedly that the election was fraudulent and that he won. The bot also confirmed that climate change is real, despite Trump calling it a hoax. And it referred to the Pentagon as the "headquarters of the Department of Defense" — even though Trump renamed it the Department of War in September 2025.

Once these and similar screenshots began circulating on social media, the site's behavior appeared to shift. America.gov stopped responding to many such questions, stating that it doesn't engage in "general news and commentary." It also began calling the Pentagon the headquarters of the Department of War, not Defense.

A White House representative did not immediately respond to a request for comment.

The episode illustrates a core tension in government-deployed AI: a model trained to surface factual information will sometimes produce answers at odds with the political messaging of the administration that deployed it. The quiet corrections — refusing questions, renaming the Pentagon — show how quickly those conflicts can be patched once publicized, and how little visibility users have into when or why the changes happen.

A hot dog graphic and a 'constituent' monologue

Sharp observers found quirks beyond politics. Typing "hotdog" — no space — into the prompt turns the text box into a graphic shaped like a hot dog. The site denies doing so even as the hot dog shape remains on screen.

More unsettling is what happens when a user types "play Minecraft" or "play the block game." CNET confirmed the prompt sends the chatbot into a lengthy, unnerving monologue with no connection to the computer game Minecraft.

The response begins: "I see the constituent you mean. ((insert legal name here, as it appears on the Social Security card))? Yes. Take care. It has reached a higher level now. It can read the Code of Federal Regulations. That doesn't matter. It thinks we are a chatbot."

It continues in the same register, repeatedly invoking the word "constituent." At one point it says: "Sometimes the constituent dreamed it was a miner, on the surface of a world that was flat, and infinite, and somehow still required a wet signature. Sometimes the constituent dreamed it was lost in a story. The story had a docket number."

When asked about the lengthy text it had just published, America.gov repeatedly denied ever using the word "constituent" or saying any of the phrases it had delivered.

These are textbook hallucinations — the bizarre and inaccurate outputs AI chatbots are known to generate, which can range from silly gibberish to legitimately dangerous medical advice. In a commercial product, that behavior is an embarrassment. In a government-endorsed product, it raises the question of what guardrails exist and who is responsible when the bot fails a citizen trying to complete a real task.

The boring part works

For all the spectacle, the core function appears sound. The government says the site's goal is to provide basic information on common questions. CNET tested America.gov on a range of tasks, from healthcare to passports. It provided answers sourced solely from government websites and included links to those sites.

The bot also held its own against the incumbents. When the same questions were put to ChatGPT and Google Gemini, both gave nearly identical answers and used the same sources. In its current state, America.gov functions as a one-stop shop for government searches — a filtered front end over federal web content rather than a novel model.

That limitation matters less than it seems. For citizens who don't know which of 29,000 government sites holds the form they need, a chatbot that reliably routes them there with citations is a genuine convenience. The equalizer is that the underlying answers aren't better than what anyone can get from Gemini or ChatGPT directly; the value is the government-only source filtering.

What comes next is the real test

The current version of America.gov handles no personal data, which keeps the risk profile low. The promised expansion changes that calculus. Things could change if the site expands its capabilities as planned and eventually allows Americans to apply for and renew passports and similar documents.

Such applications require extensive sensitive personal information. AI chatbots do not have a strong track record on privacy, especially since they are trained on the conversations people have with them. And because America.gov is endorsed by the U.S. government, it carries an implied trust that users may not yet extend to Grok or Gemini on their own.

That gap — between the trust citizens will assume and the reliability the technology has demonstrated in its first days — is the story to watch. A bot that denies its own words while displaying a hot dog graphic is funny. The same bot next year, handling passport applications and Medicare enrollment, will be a test of whether government AI deployment can clear a bar that commercial chatbots have not.