Former Meta news chief launches Forum AI to audit high-stakes model accuracy

Campbell Brown, the former dedicated news chief at Meta, has launched Forum AI to address what she describes as critical inaccuracies and biases in large language models. The New York-based company, founded 17 months ago, aims to evaluate foundation models on complex, high-stakes subjects including geopolitics, mental health, finance, and hiring. Forum AI recently raised $3 million in funding led by Lerer Hippeau, with the capital directed toward building rigorous evaluation frameworks for enterprise clients.

Brown argues that the current AI industry prioritises coding and mathematical capabilities over news and information accuracy. She contends that foundation models frequently exhibit missing context, political bias, and factual errors. In her assessment, nearly all major models display a left-leaning political bias, while specific instances, such as Google’s Gemini model sourcing content from Chinese Communist Party websites for unrelated stories, highlight systemic failures in source verification.

To counter these issues, Forum AI employs human experts to architect benchmarks and subsequently trains AI judges to evaluate model performance at scale. For its geopolitics benchmark, Brown has recruited prominent figures including Niall Ferguson, Fareed Zakaria, former US Secretary of State Tony Blinken, former House Speaker Kevin McCarthy, and Anne Neuberger. The company’s objective is to achieve a 90% consensus between its AI judges and these human experts, a threshold Brown states the company has already reached.

The venture’s business model relies on enterprise demand for liability management. Brown suggests that businesses utilising AI for credit decisions, lending, insurance, and hiring require precise, domain-specific evaluation to mitigate risk. She criticises the current compliance landscape as inadequate, noting that standardised benchmarks and checkbox audits often fail to detect violations. This was illustrated by a New York City comptroller’s finding that more than half of AI audits for hiring bias contained undetected violations.

Brown draws parallels between the current AI landscape and her tenure at Facebook, where she observed that optimising for engagement often compromised information quality. She notes that the fact-checking program she built at the social media giant no longer exists. With consumer trust in AI at extraordinarily low levels, Brown believes that rigorous, expert-led evaluation is necessary to move the industry beyond what she describes as widespread inaccuracies and towards more truthful outcomes.