Expert: David Nesting (former deputy chief information officer, U.S. Office of Personnel Management); Rumman Chowdhury (founding director, Independent AI Evaluation Foundation), Public-sector IT and AI-evaluation specialists, quoted by Axios The three underlying facts are not in dispute. Joe Biden won the 2020 presidential election with 306 electoral votes to Donald Trump's 232; a mob broke into the U.S. Capitol on 6 January 2021; and Barack Obama was born in Honolulu, Hawai'i. All three are documented in official records, so a government information service that states them one day and refuses to state them the next does not become wrong by refusing - it becomes unusable as a reference, because the user cannot tell whether the guardrail, a cached answer or the model's own sampling decided the outcome. David Nesting, the former deputy chief information officer at the Office of Personnel Management, points to ordinary mechanics rather than conspiracy. A factual question that sits close to what a model classifies as "too political" can tip either way "purely probabilistically", and the two model families behind the service, Gemini and Grok, may treat the same question differently. America.gov also states that some answers may be cached for up to two hours, so a user can receive a stale answer even after behaviour has changed. Rumman Chowdhury, founding director of the Independent AI Evaluation Foundation, accepts the probabilistic explanation but warns that guardrails could "also be shaping the output to be ideologically aligned with what the president or his party wants". The problem for readers is that none of these explanations is verifiable from outside. The GSA did not answer Axios' questions about America.gov's guardrails, and the assistant told Axios it "cannot share or reconstruct internal instructions". On an ordinary chatbot that opacity is a quality issue; here the answers carry the authority of a federal service. The stakes rise with the next release planned for early 2027, which will let users complete forms, applications and renewals - including Medicare enrolment - and will require them to hand over personal information to a system Nesting worries will become "one of the largest systems holding very, very confidential private information". Good AI practice for a service at this level means published guardrails, versioned answers and a visible change log, so that a fluctuation can be told apart from an override. Source: https://www.axios.com/2026/09/30/trump-america-gov-ai-chatbot-conflicting-answers