Particle.news

Study Finds Major Chatbots Often Wrong on Elections and Cite State Media

The results have prompted demands for independent audits and greater transparency as chatbots take on more news queries ahead of the US midterms.

Overview

  • Wednesday's Forum AI study asked four chatbots more than 3,100 news questions and found collective failures on accuracy, bias, or source selection in most election-related answers.
  • The report says nearly 36% of answers to election questions contained at least one factual error and that xAI's Grok returned errors in roughly half of its election responses.
  • Researchers found frequent reliance on foreign state-controlled outlets in foreign-policy answers, with ChatGPT and Grok citing such sources about half and 44% of the time respectively.
  • The study flagged a calibration problem where answers look confident and professionally cited yet contain hidden factual mistakes, which Forum AI says raises risks for voters who use chatbots for news.
  • Anthropic invited reviewers to examine the data and defended Claude's neutrality while Forum AI and others urged third-party audits and clearer source transparency ahead of the midterm cycle.