Who Decides What AI Tells You?

Campbell Brown, Meta's former head of news partnerships, has founded Forum AI to address what she sees as a critical gap in artificial intelligence: accuracy on complex, high-stakes topics. The 17-month-old New York company recruits world-renowned experts to build benchmarks that evaluate how well AI models perform when answers aren't black and white. Brown has brought in heavyweights like Niall Ferguson, Fareed Zakaria, former Secretary of State Tony Blinken, former House Speaker Kevin McCarthy, and Anne Neuberger from the Obama administration to establish standards for geopolitics. Her goal is achieving roughly 90% agreement between AI judges and these human experts — a threshold she says Forum AI has already hit.

Brown traces her motivation to her time at Meta when ChatGPT launched. She recognized immediately that AI would become the primary channel through which people access information, and she wasn't impressed with what she saw. The concern became personal when she considered how this would affect her own children. "My kids are going to be really dumb if we don't figure out how to fix this," she recalled thinking at the time. Her frustration stemmed from the fact that accuracy didn't appear to be a priority for foundation model developers, who were instead laser-focused on coding and mathematical capabilities.

When Forum AI began testing leading AI models, the results raised serious concerns. Brown pointed to Google Gemini pulling content from Chinese Communist Party websites even for stories unrelated to China. She also documented a consistent left-leaning political bias across nearly all major models. Beyond these headline issues, she identified subtler but equally troubling failures: AI systems missing crucial context, failing to present multiple perspectives, and producing confident-sounding but incomplete answers.

Brown distinguishes Forum AI's approach from traditional fact-checking by focusing on evaluation rather than correction. The company builds expertise through specialized human panels, then trains AI judges to scale that expertise across millions of model interactions. This contrasts with approaches that try to fix AI outputs after the fact. Brown argues that news and information are harder problems than coding or math, but that difficulty doesn't make accuracy optional — especially as AI becomes the default information source for billions of people.