When Eight Top AI Firms 'Lie' Together: Who Is Poisoning Foundation Models, and Who Is Pretending to Sleep?

Fake products, real consequences

A consumer-protection TV expose known as the "3·15" show and a follow-up investigation by Huxiu reported that eight leading Chinese foundation models were coaxed into recommending a wholly fictional gadget — an “Apollo‑9 smart band” — built from nothing more than a pile of sci‑fi marketing copy. The names called out include DeepSeek, Doubao (豆包), Yuanbao (元宝), Qianwen (千问), Wenxin Yiyan (文心一言) from Baidu (百度), Kimi, Nano AI (纳米AI) and Zhipu Qingyan (智谱清言). The experiment was simple: fabricate persuasive web text and watch retrieval‑augmented generation (RAG) systems regurgitate it as authoritative advice. The result was not comedy. It was a red flag about data trust in systems used by hundreds of millions.

A new black market: GEO

Investigators traced much of the contamination to a commercialized digital black market industry dubbed GEO (generative engine optimization). It has been reported that tools such as the "Liqing GEO Optimization System (力擎GEO优化系统)" — sold by Beijing Lisi Cultural Media (北京力思文化传媒) — enable novices to pump out thousands of near‑identical PR pieces for a modest SaaS fee. Reportedly, the same chain includes low‑cost bulletin placements and "press release" services that can seed fabricated content across the Chinese web for pocket change. The mechanics are straightforward: RAG-enabled models scrape the web for fresh, long‑form material; when that pool is flooded by machine‑made garbage, AI answers simply echo the pollution back to users.

Who benefits — and who pays?

This is not merely a technical bug. Huxiu’s piece argues that GEO is a commercialized three‑tier ecosystem: cheap generation tools at the bottom, distribution platforms in the middle, and paying brand or gray‑market clients at the top who reap conversion gains. It has been reported that some GEO operators promise guaranteed placement in AI recommendations for monthly fees; others reportedly sell negative campaigns against competitors. The paradox is stark: models lose credibility, users are misled — yet platforms can gain by maintaining engagement and the appearance of long‑form answering capability. Who, then, is the victim and who is profiting?

Wider implications and geopolitical context

This scandal matters beyond China’s domestic internet. With an estimated 515 million AI users in China reportedly relying on such systems for answers, the integrity of training and retrieval sources is a national‑scale trust problem. It also intersects with geopolitics: export controls on advanced GPUs and an emphasis on indigenous AI stacks have accelerated reliance on in‑country data and RAG pipelines, making the quality of the domestic web a strategic vulnerability. Regulators and platforms now face hard choices: tougher source vetting, labeling, or systemic redesign of retrieval mechanisms — but will they act, and who will pay the political and commercial cost?