EN ▾
SNAPE REP RT Front page →
ConfirmedAI safety04 Nov, 11:09 UTC

Researchers' test sites about real Senate candidates were cited by chatbots within 19 hours, Caucus AI reports

A voter at the polls during the midterm elections in Brooklyn, New York
Leonardo Munoz / AFP via Getty Images, via NBC News

In its post-election report the independent project says about a third of answers linking its sites called them official candidate sites, and that ChatGPT, Gemini and Google's AI Overviews discussed a fictional election as if it were real.

Read at NBC News Share

Why it matters

The test shows how cheaply outsiders can shape what chatbots tell voters: the sites were generated with an AI agent, cost about $12 per domain and took about 48 hours to launch, with disclaimers saying they were not official. A few sentences on data centres made answers about those candidates 4.4 percentage points more likely to mention data centres. The report also finds 44% of likely voters use AI products at least weekly. Watch whether OpenAI and Google change how their search tools rank new sites.

Key facts

  • Test sites cited within 19 hours
  • About a third of linking answers called them official sites
  • Sites launched on 10 October; about $12 per domain
  • Data-centre text raised mentions by 4.4 percentage points
  • 44% of likely voters use AI products at least weekly

Sources

More on AI safety
More from Snape ReportFront page →