Campbell Brown, a veteran journalist who once ran CNN's news desk and later Meta's global news policy, spent 17 months after leaving Meta doing something most foundation model companies skip: systematically evaluating the information quality of mainstream AI outputs. Her findings are stark — nearly every model tested shows a structural left-wing political bias, and Google's Gemini even cited content from the Chinese Communist Party's official website in reports completely unrelated to China.
Gemini's Source Selection Glitch: Context Invisible
Brown founded Forum AI in New York, creating a geopolitical event benchmark framework. She assembled a panel of advisors with diverse political and professional backgrounds — including Niall Ferguson, Fareed Zakaria, former Secretary of State Tony Blinken, former House Minority Leader Kevin McCarthy — to rate AI responses. The system now aligns with human expert consensus roughly 90% of the time.
Her findings break down into three layers, each harder to fix than the last. Layer one: source selection logic failures. Gemini, when handling reporting unrelated to China, pulled text from the CCP's official website. This isn't a fact error in the usual sense — it's a sign that the model's crawling logic doesn't evaluate source credibility or political agenda. The political nature of a source remains invisible in the AI's output pipeline.
Structural Left Bias: A Training Data Artifact
Layer two: structural political bias. Brown found nearly all mainstream models exhibit left-leaning tendencies. This isn't conspiracy — it's a natural result of training data distribution. AI mimics the tone and framing of the texts it learns from. English-language internet content — mainstream news, academic papers, social media — carries certain political leanings, and models inherit those biases without knowing they're doing so. This isn't a bug you patch; it's embedded in every output logic.
Missing Multi-Perspective: AI Delivers Statements, Not Debates
Layer three: lack of context and multiple viewpoints. Brown says current models universally lack "context, multi-perspective, and argument transparency". They give a statement, not a structure like "here's what side A believes, here's what side B believes, here's their fundamental disagreement." They offer an answer without disclosing which angle it comes from.
Brown points to a structural blind spot: foundation model companies prioritize math, coding, and logic in their benchmarks. Information accuracy and political diversity rarely appear on mainstream leaderboards. Code has right and wrong answers; math has standard scores. But what constitutes an accurate and fair geopolitical news report? Who decides? How many people from how many perspectives need to agree? There's no engineering solution, so the question gets systematically skipped. The result: information accuracy is virtually an invisible metric.
The cost of skipping it appears in a concrete case: New York City's audit of AI hiring tools found that over half the cases flagged no violations. Brown argues this may not mean low violation rates — it may mean the auditing AI itself lacks accuracy to detect problems, not that problems don't exist.
Fluently Wrong: Harder to Catch Than Silence
Brown's core thesis: AI's problem isn't just making factual errors — it's getting people to trust and accept those errors. When a model delivers a wrong answer in a smooth, confident, unhesitating tone, most users have no reason to doubt it. Fluently wrong is harder to catch than silence, and harder to correct.
What will drive change? Brown is blunt: not moral pressure or public opinion, but corporate compliance risk. Under current incentive structures, no one has a strong enough reason to fix this until the cost becomes unavoidable. Credit approval, insurance underwriting, hiring screening — these AI decisions are already subject to anti-discrimination laws. Once an AI output produces discriminatory or inaccurate results, the business using it bears liability. That pressure will push back to model providers, requiring auditable, verifiable outputs — not because it's morally right, but because enterprise contracts start demanding it.
Lerer Hippeau led Forum AI's $3 million seed round last year — a small sum in AI, but a signal that "AI evaluation" is a business whose demand may grow faster than it currently appears.