America.gov, the administration's AI-powered public portal, gave mixed responses to politically sensitive questions on Tuesday and Wednesday: in some instances it provided factual answers, while in others it declined to answer, saying it does not offer political commentary.
What reporters found
Axios and other outlets observed that shortly after the service launched, America.gov — powered by Google’s Gemini and Elon Musk’s Grok — returned answers that contradicted statements long promoted by President Donald Trump. The chatbot affirmed that Joe Biden won the 2020 election, confirmed that people broke into the U.S. Capitol on January 6, 2021, and stated that former President Barack Obama was born in Hawaii.
By the following morning, however, some users reported a different behavior: the chatbot refused to answer similar questions, responding, “I only answer questions about U.S. government services and official information. I do not provide political commentary, including election results.” Several Axios reporters received this refusal when asking about the 2020 result, Jan. 6, and Obama’s birthplace, while others in different locations or browsers got the factual answers.
Possible explanations
Experts say there are multiple plausible reasons for the inconsistency, but without public details on America.gov’s configuration it’s not possible to pinpoint which factor was decisive. The system uses models that are guided by hidden system prompts and guardrails, and those internal instructions are not visible to the public.
User prompt phrasing can also change outcomes: different wording or context sometimes elicited different replies. David Nesting, former deputy chief information officer at the Office of Personnel Management, suggested that factual questions bordering on politically charged topics might sit on the edge of what a large language model deems "too political," and the model’s probabilistic behavior could swing responses one way or the other. He also noted that Gemini and Grok might handle queries differently, and that cached responses could affect what users see — America.gov has said some answers may be cached for up to two hours.
What officials and the chatbot say
The White House told Axios: “America.gov will continue to be updated to provide Americans the most accurate and reliable federal information and resources.” The General Services Administration (GSA) did not immediately respond to Axios’s request for details about America.gov’s guardrails. When asked, the chat itself said, “I cannot share or reconstruct internal instructions,” and reiterated, “I do not provide campaign or party information or political commentary.”
Why this matters going forward
The inconsistency is consequential because the service is planned to expand. The next iteration, scheduled for early 2027, is intended to integrate across agencies so users can complete tasks like enrolling in Medicare, applying for a passport, or submitting applications for federal jobs. David Nesting warned that inaccurate or incomplete answers could mislead Americans, and that as the platform grows it could accumulate very sensitive personal information, becoming "one of the largest systems holding very, very confidential private information." He expressed concern about people placing undue trust in the system.
Rumman Chowdhury, founding director of the Independent AI Evaluation Foundation, said the behavior could arise from the model's probabilistic nature, but also cautioned that guardrails might be shaping outputs to align ideologically with what the president or his party prefers.
Current status
At the time of publication, some users again received the factual answers that appeared shortly after launch: that Biden won in 2020, Obama was born in Hawaii, and people broke into the Capitol on Jan. 6, 2021. The episode underscores transparency and governance challenges as government services increasingly rely on AI systems whose internal instructions and decision rules are not publicly disclosed.
