Who decides what AI tells you? Campbell Brown, once Meta’s news chief, has thoughts

A New Approach to AI Accuracy

Campbell Brown, who previously led news partnerships at Meta, is now tackling AI misinformation with her new company, Forum AI. The startup aims to evaluate how foundation models perform on complex, "high-stakes" subjects where accuracy is critical but difficult to measure, such as geopolitics and finance. This move addresses a growing concern that model developers are prioritizing capabilities like coding over the nuanced understanding required for reliable information delivery.

The Expert-Led Framework

Forum AI, which raised $3 million in a round led by Lerer Hippeau, has developed a distinct evaluation methodology. Instead of relying on standardized tests, the company recruits world-renowned experts—including figures like Niall Ferguson, Fareed Zakaria, and former Secretary of State Tony Blinken—to design sophisticated benchmarks. Early evaluations have already revealed significant model flaws, Brown said, from pervasive political bias to citing Chinese state propaganda on unrelated topics.

• Expert-Driven Benchmarks: Domain leaders architect tests for nuanced, complex subjects.
• AI Judge Training: AI evaluators are trained to replicate expert consensus on model performance.
• Performance Target: The goal is to achieve approximately 90% consensus between the AI judges and human experts.

The Enterprise Push for Reliability

Brown believes the primary driver for improved AI accuracy will not be consumer demand, but enterprise necessity. Businesses using AI for liability-sensitive applications like hiring, credit decisions, and insurance cannot afford the "slop and wrong answers" common in consumer-facing chatbots. By providing a more rigorous auditing process than what Brown calls the current "joke" of compliance checks, Forum AI is betting that corporate liability concerns will create a durable market for high-fidelity model evaluation, ultimately forcing the industry to prioritize correctness.

While foundation model developers chase performance on coding and math benchmarks, Forum AI is betting that enterprise liability will be the real forcing function for accuracy in high-stakes domains, creating a new market for expert-led, nuanced evaluation that current standardized tests miss.