America.gov gave factual answers about the 2020 election and Jan. 6, then began refusing some similar questions as “political commentary.” The White House says the system is still being updated.
The federal government has a new AI chatbot, and within days it managed to stumble into the one assignment guaranteed to produce chaos in Washington: answering political questions.
America.gov was built as an AI-powered front door to federal information, capable of pulling material from tens of thousands of government webpages. Users quickly began asking it questions about the 2020 election, January 6 and other politically sensitive topics. At first, the bot gave straightforward factual answers.
It said Joe Biden won the 2020 election. It described people breaking into the U.S. Capitol on January 6, 2021. It confirmed basic historical facts that conflict with false claims Trump has repeatedly promoted.
Then some users started getting a very different response. Similar questions were suddenly rejected as requests for “political commentary,” while other users still received answers depending on the wording or timing of the prompt.
That makes the chatbot’s behavior more interesting than any one answer. There is no evidence that the bot developed an anti-Trump personality, joined the resistance or quietly opened a Bluesky account. The likely explanations are much less dramatic: prompt sensitivity, model variability or developers changing political guardrails behind the scenes.
The government has not fully explained which of those things happened. Axios reported that the General Services Administration said the system is still evolving and being updated.
That would be a fairly ordinary disclaimer for a consumer AI product. It becomes more complicated when the product is an official federal information service that may eventually help people access Medicare, apply for government jobs or complete other federal tasks.
An inconsistent answer about politics can be funny. An inconsistent answer about eligibility for a federal benefit would be considerably less entertaining.
The political questions highlight a deeper design problem. Government agencies are supposed to communicate stable information. Generative AI is probabilistic by nature, meaning two users can sometimes receive different answers to nearly identical prompts. Add political filters on top of that, and the system can start looking less like a reference desk and more like a bureaucrat trying to guess whether the question is going to get anyone in trouble.
The original answers were not inherently partisan. Saying Biden won the 2020 election is a statement of certified election results. Describing the Capitol breach is a matter of public record. The controversy begins when those questions are treated as political commentary rather than factual history.
That distinction matters because official government systems should ideally separate facts from opinions rather than avoid facts merely because politicians argue about them.
The White House says America.gov remains under development, and that is a reasonable explanation for early inconsistency. AI products often change rapidly after launch as developers adjust filters and fix unexpected behavior.
But users have no clear way to know when a refusal reflects a safety rule, a political-content rule or a quiet policy change made after somebody noticed a screenshot circulating online.
That is the weirdest part of the whole episode. The government built an AI tool to make federal information easier to access, then immediately demonstrated why transparency around the AI itself may be just as important as the information it provides.
So the question is not whether America.gov should have political opinions. It should not. The question is whether an official government chatbot should ever refuse to state basic facts simply because those facts happen to be politically inconvenient.
