Descriptions:
Forum AI CEO Campbell Brown joins Bloomberg Technology to present findings from NewsBench Wide, an independent benchmark evaluating major AI chatbots on news accuracy, political bias, and source quality. The headline result: chatbot responses to election-related questions failed on at least one of accuracy, bias, or source selection 90% of the time across four major platforms — Claude, Gemini, ChatGPT, and Grok.
The political bias findings are particularly striking. Claude and Gemini produced left-leaning answers on election questions 100% of the time in Forum AI’s testing; ChatGPT did so 95% of the time; Grok was the outlier, giving right-leaning responses approximately 85% of the time. Beyond bias, about one-third of all election-related answers contained outright factual errors, and multiple chatbots cited Russian and Chinese state media as authoritative sources.
Brown, a former CNN and NBC News anchor and ex-Meta VP of News and Global Media Partnerships, argues the core problem is structural: AI companies are currently grading their own homework, with no independent evaluation ecosystem for news accuracy the way one exists for coding and math benchmarks. She calls for a mix of private companies and nonprofits to provide third-party evaluation, noting that several U.S. states are already moving toward legislative requirements for independent assessment. Gemini showed the most improvement potential across tested models. The conversation is directly relevant to any enterprise deploying AI for information-sensitive use cases ahead of the 2026 midterm elections.
📺 Source: Bloomberg Technology · Published May 20, 2026
🏷️ Format: Benchmark Test
