America.gov Puts 29,000 Government Sites Behind One AI Chatbot

The U.S. government has launched an AI chatbot at America.gov, built with help from Google and SpaceXAI, according to Sept. 29 reporting from TechCrunch.

The site describes the system as using AI to provide simple answers only from official government sources. Every answer is sourced exclusively from federal, state and local websites, and the project is presented as bringing together 29,000 government websites into one place, according to America.gov.

Access is browser based. No app download is required. The service works on phone, tablet or computer. It is free to use and contains no advertisements, according to America.gov.

The privacy description is brief. America.gov states it does not collect or store personal information and that conversations disappear when users leave, according to America.gov. The scope is broad. The limit is narrow by design, at least as described: official government domains only.

More transactional functions are described as a plan, not a current capability. America.gov plans to let users complete forms, track progress and organize everything in one place starting in 2027, according to America.gov.

Staffing around the build has a named lead. Donald Trump said 20-year-old programmer Edward Coristine was a lead engineer on the America.gov project, according to TechCrunch. Coristine is known by the nickname "Big Balls" and was involved in Elon Musk's DOGE, according to the same reporting.

Two early behaviors have drawn attention. The chatbot states that Joe Biden won the 2020 election. Asking it about Minecraft produces a roughly 1,800-word monologue, according to TechCrunch. That response is a rewriting of the End Poem written by Julian Gough that appears after beating Minecraft.

Looking at what this means for practitioners, the core technical tension will be familiar. A general-purpose model that can generate open-ended text is being asked to act like a closed lookup tool. Limiting it to an allowlist, a fixed list of approved government domains, lowers the risk of hallucination, when a model invents facts, and makes the source easier to trace. It does not remove prompt sensitivity, where small wording changes alter the answer, shifts in tone, or edge cases where the model fills gaps with fluent text.

In my view, the Minecraft example is more instructive than embarrassing. A long, lyrical answer to a topic outside its official corpus suggests system instructions that favor answering over refusing, or a fallback path that was not tightly scoped. For civic infrastructure, that calibration matters as much as corpus size. Citizens will test boundaries. They will ask about benefits eligibility in one turn and video games in the next. Consistent refusal behavior, citation granularity, versioning of sources, and audit logs are where trust will be won or lost.

The broader context here is constructive. A single conversational entry point across thousands of fragmented federal, state and local sites addresses a real discovery problem, much like a single front desk for scattered offices. Forms, status tracking and persistent organization, if delivered as described for 2027, would move the project from question answering toward task completion. That is the harder engineering task, involving identity, entitlements, and cross-agency data flows, but it is also where time savings compound.