Former Meta News Chief Investigates: Nearly All AI Models Lean Politically Left, Gemini Once Cited Chinese Communist Party State Media

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯2196 words

Campbell Brown, former head of news at Meta, founded Forum AI and spent 17 months systematically evaluating the information quality of mainstream AI models. She discovered that Gemini had cited information from Chinese Communist Party official websites when handling reports unrelated to China, and that nearly all tested models exhibited left-leaning political bias. (Background: AI destroys 133-year Princeton University tradition: When "cheating" gradually becomes common sense) (Context: 97,895 underground forum conversations tell you: the hacker community actually hates AI too) Brown started as a journalist and once served as an anchor at CNN, later transitioning to head of news at Meta, where she directly managed the policies governing how Facebook presented news to 3 billion users worldwide. This position gave her a close-up view of the full picture of "how platforms shape information flow." She left Meta 17 months ago and founded Forum AI in New York, dedicated to doing one thing that foundation model companies generally skip: systematically evaluating whether the information AI provides is accurate, fair, and offers diverse perspectives. Forum AI's core product is a "geopolitical events benchmark framework." Here's how it works: Forum AI invites a group of top advisors with different political spectrums and professional backgrounds — Niall Ferguson, Fareed Zakaria, former U.S. Secretary of State Tony Blinken, former House Minority Leader Kevin McCarthy, former U.S. Deputy National Security Advisor Anne Neuberger — to individually score the answers from mainstream AI models on the same complex geopolitical event. Currently, Forum AI has reached a threshold of approximately 90% consensus with human experts, making Forum AI's evaluation results a defensible benchmark rather than just one person's opinion. The problems Brown discovered fall into three layers, each more difficult to fix from a technical standpoint than the last. The first layer is flaws in source selection logic. Gemini, when handling certain reports unrelated to China, cited content from Chinese Communist Party official websites. This isn't a factual error in the conventional sense, but rather a problem with the model's filtering logic when scraping sources: AI only judges "this is text, this is a link," not "what is this source's stance, how credible is it, does it have a clear political purpose." The political nature of sources themselves is invisible in AI's output process. The second layer is structural political bias. Nearly all mainstream models Brown tested exhibited left-leaning political bias. This isn't a conspiracy theory, but a natural result of training corpus distribution. AI tends to replicate the tone and stance framework of whatever text it learns from. The mainstream content of the English internet — mainstream media reporting, academic papers, social media posts — overall carries specific political leanings, and the models trained on it inherit this tendency, without being aware that it is doing so. What's even trickier is that this bias isn't a bug that can be found and patched, but is embedded in every output logic of the model. The third layer is the lack of context and multiple perspectives. Brown said existing models generally lack "background context, multiple perspectives, and argumentative transparency." The answers AI provides are declarative sentences, not structured as "from faction A's perspective this means this, from faction B's perspective it means that, and the fundamental disagreement between them lies in…" It gives you an answer, but doesn't tell you from which angle that answer originates. Brown pointed out a structural blind spot: when foundation model companies evaluate and rank models, their priorities are mathematics, coding, and logical reasoning capabilities. Information accuracy and political diversity almost never appear on mainstream benchmark testing lists. The reason isn't hard to understand. Code has right and wrong answers — run the tests and you'll know. Math problems have standard answers, and accuracy can be calculated. But "what constitutes accurate and fair reporting on a geopolitical news story" — who decides? How many people with different stances need to form a consensus before it counts? This question has no engineering solution. In a product development process led by engineers and where market positioning is determined by benchmark rankings, it gets systematically skipped. As a result, information accuracy is an almost invisible metric in AI's capability evaluation system. The cost of being skipped can be seen in a concrete case. New York City last year conducted a round of compliance audits on AI hiring systems, aimed at checking whether AI screening tools used by employers violated current anti-discrimination employment regulations. According to the audit results, over half of the cases failed to detect violations. The problem with this number isn't "low violation rate," but rather what it might represent: that the AI tools performing the audits themselves lack sufficient accuracy, to the point that they cannot see where the problem lies, rather than the problem not actually existing. This is the core of Brown's argument: AI's problem isn't just providing wrong facts, but causing people to accept wrong facts with trust. A person who knows they don't know something at least has the opportunity to look it up. But when AI delivers a wrong answer in a fluent, confident tone without hesitation, most users have no reason to doubt it at all. Fluent errors are harder to detect and harder to correct than silence. Brown's judgment is straightforward: what will drive change won't be moral pressure or public opinion, but the commercial pressure brought by corporate compliance risk. Behind Brown's argument lies a realist foundation: under the existing incentive structure of the AI industry, no one has a sufficiently strong reason to actively solve this problem, until its cost becomes impossible to ignore. Credit approval, insurance underwriting, hiring screening — AI decisions in these scenarios are all subject to current regulations. Once AI outputs discriminatory or inaccurate results, the businesses using AI bear legal liability. This pressure will ultimately transmit upward to model providers, demanding that they deliver auditable, verifiable outputs with guaranteed accuracy. Not because they think it's morally correct to do so, but because enterprise client contracts have begun to write this requirement in. Lerer Hippeau led Forum AI's $3 million seed round last year. This figure is small money in the AI field, but what it represents is a judgment: "AI evaluation" is a business, and the demand for this business may grow faster than is currently visible.

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-30

Meta launches "AI necklace" test, aiming to ship 10 million wearable devices in the second half of the year

Similarity 120%關鍵字 meta同分類 zh

2026-05-29

Gemini Prediction Market Upgrades AI Brain: Integrates Grok Customization, Automatically Recommends Your Watchlist

Similarity 120%關鍵字 gemini同分類 zh

2026-05-28

Meta expands subscriptions: IG, FB, and WhatsApp launch Plus plans, allowing anonymous viewing of others' Stories

Similarity 120%關鍵字 meta同分類 zh

2026-05-26