Context:

America.gov’s AI chatbot, powered by Google Gemini and Grok, gave mixed responses to politically sensitive prompts about the 2020 election, the Jan. 6 attack, and Barack Obama’s birthplace, with some users seeing answers and others being told the bot only provides government information. The inconsistencies appeared to vary by user, device, and framing, prompting questions about how the system is configured and what it is allowed to say. Officials noted the lack of transparency around guardrails, while experts suggest outputs may shift due to system prompts, user prompts, or the model’s political guardrails. The episode comes as the service prepares a major 2027 update to expand federal features such as Medicare enrollment and passport applications, though the GSA has not commented on guardrails.

Dive Deeper:

• America.gov, an AI-driven government information service, leverages Google’s Gemini and Elon Musk’s Grok to help users access federal information and services.

• Early reports showed the chatbot contradicting President Trump on the 2020 election and January 6, and even confirming Obama was born in Hawaii, countering long-standing conspiracy theories.

• By the next day, some users encountered refusals to answer, with the bot saying it does not provide political commentary or election results, while other users still received substantive answers depending on context.

• Experts cite multiple possible causes for variance, including system prompts, user framing, and how instructions are integrated into the model, though exact configurations remain opaque.

• There is concern that outputs could be ideologically influenced by administration direction, compounded by a lack of transparency about guardrails and internal setup.

• The service is slated for a major update in early 2027 to incorporate additional federal services like Medicare enrollment and passport applications, but the General Services Administration has not publicly addressed guardrail details.