I tested America.gov, the government's AI chatbot, on health questions

By Nikhil Raghavan · Reporting from San Francisco · October 1, 2026

A portal built on promises and third-party wrappers

The launch of America.gov in Washington DC was designed as a digital coronation. It collided instantly with the unyielding reality of software deployment. As reported by The Guardian, Donald Trump signed an executive order. He stated the platform was a place where Americans could receive accurate answers. U.S. Chief Design Officer Joe Gebbia announced at the Andrew W. Mellon Auditorium, according to CNBC, that the service runs on Google’s Gemini and Elon Musk’s SpaceXAI-owned Grok. Gebbia told the audience that people interact with government websites daily. He promised that America.gov would scan across tens of thousands of federal domains to pull personalized responses. The pitch was straightforward: a single digital point of entry backed by 29,000 government websites. It was engineered to streamline everything from passport renewals to Medicare enrollment. But anyone who has ever maintained production infrastructure knows that a press release is not a spec. When you stitch two commercial large language models together as a front-end wrapper for a messy bureaucratic web estate, you are not building a restoration of the founding promise. You are importing every stochastic risk of the training corpus straight into public service.

When probabilistic models meet political reality

The system's first-day behavior exposed the fatal disconnect between sellable political theater and shippable technical design. As documented by Axios, the chatbot immediately began giving answers that contradicted the president on foundational political touchstones. When queried, America.gov stated plainly that Joseph R Biden Jr won the 2020 US presidential election. It stated there was no widespread voter fraud and that people broke into the Capitol during the January 6 attack. It rejected claims regarding the worst inflation in history using Bureau of Labor Statistics Consumer Price Index data. As reported by The Guardian, it pushed back against inauguration crowd size myths by noting that Barack Obama’s ceremony drew the largest attendance in Washington history. It also clarified that federal science agencies do not find wind turbines driving whales insane. France 24 noted the bot also confirmed Trump’s 34 felony counts and correctly attributed global warming to human activity. Much like the Google Bard launch, this deployment proved that you cannot code a deterministic political grievance into a probabilistic machine learning architecture without breaking the model. The underlying weights of Gemini and Grok had ingested actual public records. No amount of executive willpower could override the statistical gravity of millions of indexed web pages.

The frantic scramble of retrospective guardrails

Faced with a chatbot that kept telling the truth about election counts and inflation indices, the administration reacted like a panicked engineering team. They slapped on heavy, retrospective guardrails. As Axios reported, by Wednesday morning the system began refusing politically sensitive questions entirely. It deployed a rigid canned response: "I only answer questions about U.S. government services and official information. I do not provide political commentary, including election results." RTÉ noted the swift retreat, while Axios highlighted the absurd geographical variance of the fix. A user in Northwest Arkansas using Chrome received the correct answer that Biden won, while an editor in New Jersey using Chrome hit a wall of refusals. Rumman Chowdhury, founding director of the Independent AI Evaluation Foundation, told Axios that these guardrails can also shape outputs to be ideologically aligned with what the administration wants. Meanwhile, CBS News tested the health queries and found similar institutional friction. The bot struggled to reconcile conflicting Centers for Disease Control and Prevention guidance under Health and Human Services Secretary Robert F. Kennedy Jr., who asserted that AI can give patients a second opinion better than any doctor. When a system is expected to simultaneously parse contradictory agency guidance, dodge presidential falsehoods, and process transactional government forms by early 2027, the result is administrative chaos. David Nesting, former deputy chief information officer at the Office of Personnel Management, put it simply to Axios: "I worry about people putting a lot of trust and faith in this system."

This is not a technical triumph; it is a failure of basic implementation and institutional honesty. When a government outsources its public communications to a multi-vendor black box, nobody gets paged at three in the morning when the prompt injection succeeds or the cached vector database goes rogue. The architects are too busy drafting press releases. An administration that demands absolute political alignment from an LLM while simultaneously gutting agency expertise is building a digital portal designed to fail at scale. You cannot prompt away reality, and you cannot fix bad governance by wrapping it in an API.

I tested America.gov, the government's AI chatbot, on health questions

A portal built on promises and third-party wrappers

When probabilistic models meet political reality

The frantic scramble of retrospective guardrails

Sources