On Friday, ByteDance’s Doubao — China’s biggest consumer AI app — told users its custom agent feature goes offline July 15, citing “product function adjustments.” The personas people built — named companions, tutors, role-play characters with fixed personalities — stop working that day, and the data behind them becomes unrecoverable after October 15. On Saturday, Alibaba’s Qwen went further: its humanlike and user-created agents shut down this Friday, July 10, with broader agent functions gone by the 15th, chat histories included. Tencent quietly pulled the same feature from its Yuanbao assistant in June. Per the South China Morning Post and Chinese state media, none of this is product strategy. It’s compliance.
The deadline is the Interim Measures for AI Anthropomorphic Interaction Services — issued in April by the Cyberspace Administration of China and four other agencies, effective July 15. It is the world’s first national law aimed squarely at AI that simulates human personality, thinking patterns, and communication styles for “sustained emotional interaction.” The rules ban virtual partners and virtual family members for minors, require guardian consent for users under 14, and mandate anti-addiction systems and identity checks. Customer-service bots, knowledge Q&A, workplace assistants, and education tools are exempt — as long as they don’t get emotional. The scale is not niche: QuestMobile counted 345 million monthly Doubao users in March, more than 200 million of them daily, with Qwen at 166 million — over half a billion users between the two apps now watching features disappear.
Beijing is not anti-agent — that’s what makes this sharp. In May, regulators issued guidance encouraging the “managed development” of AI agents; in June, China published national standards for agent identity, discovery, and tool use, days before it open-sourced two frontier models. “Current agents are not yet mature,” MIIT expert committee member Pan Helin told the SCMP, framing the policy as safety plus standardization. The message to industry is unambiguous: agents that do work get infrastructure; agents that form bonds get regulated out. Users noticed — Weibo filled with people mourning companions they’d talked to for years, with no way to export the conversations.
Our take: A category serving half a billion users just got legislated to zero in nine days’ notice — and the operative word is deleted, not discontinued. Every AI relationship lives on someone else’s server with a kill date the user doesn’t control; China just made that visible at scale. For builders and investors, emotional AI is now provably the most fragile layer of the stack — the US is circling the same territory via lawsuits and teen-safety scrutiny rather than a five-agency decree, but the direction rhymes. Meanwhile no one can guarantee an agent obeys, and the week’s Geneva dialogue shows how far the rest of the world is from writing anything binding. China regulated first and let the products catch up. Watch who copies the template.
What to watch
- Friday, July 10. Qwen’s custom agents go dark — the first hard cutoff, five days before the law itself.
- July 15. The Measures take effect and Doubao’s agents follow. First enforcement actions will show whether role-play games and virtual idols get swept in too.
- October 15. Doubao’s deleted-agent data becomes permanently unrecoverable — the quiet deadline that ends any rollback hope.
- Washington. Reports in early July put the White House in talks with top US labs on voluntary frontier-model standards. Whether companion AI makes it into that framework is the tell for US platforms.