Tests: chatbots rely on three obscure AI-built voter sites for US midterm answers; Copilot cited the error-prone Ballot Brief in 75% of tests and ChatGPT in 22%

Chatbots may cite campaign advertising as factual news, potentially misleading voters.

A network of 37 websites designed to look like local news outlets, but which function as thinly veiled political advertisements for Democratic candidates, has been cited by OpenAI's ChatGPT in approximately half of test queries. The sites, whose origins are unknown, promote specific Democratic House candidates, often with limited mention of their opponents. Ten of these sites pose as dedicated local news, each promoting one Democrat in a competitive race. A newsroom analysis revealed that these sites are part of a larger operation including roundup-style national sites and state-focused sites. Google's search and AI products did not surface these sites in their results, and the company stated it took action after detecting a policy violation. OpenAI acknowledged that its tools cite from a range of sources.

The discovery highlights a new vector for political advertising, where AI models can inadvertently amplify undisclosed campaign messaging by citing obscure, AI-optimized websites. This network's deliberate structure, including `llms.txt` files, suggests a design intended to be easily consumed by AI crawlers, bypassing traditional content moderation and editorial scrutiny that might otherwise flag such sources. The reliance on these sites, particularly in downballot races with sparse credible reporting, points to a vulnerability in how AI synthesizes information for users seeking election guidance.

This situation raises concerns for both AI developers and voters. Developers must enhance their models' ability to discern credible news from disguised political content, especially when traditional sources are scarce. For voters, it underscores the critical need for skepticism and cross-verification of information obtained through AI, as the lines between genuine local news and partisan propaganda become increasingly blurred and automated.