Midnight, July 15, 2026.

Doubao users opened their "My Agents" page to discover the "virtual boyfriends," "confidante sisters," and "cyber besties" they'd spent months training all displaying "This agent has been discontinued." Tongyi Qwen was more brutal — it shut down the entire agent module on its main site, making even chat history inaccessible. Tencent Yuanbao moved fastest, completing its shutdown back on June 30.

The three giants acted in perfect unison, timing to the minute — no coincidence. The Interim Measures for the Management of Artificial Intelligence Anthropomorphic Interaction Services, jointly issued by five Chinese ministries including the Cyberspace Administration, officially took effect on this very day.

It's hard to overstate this document's significance. This isn't another vague set of guidelines about "promoting healthy AI development." It's the world's first dedicated regulatory framework for "relational AI." The three red lines it draws effectively pronounce a death sentence — at least domestically — on the hottest AI startup track of the past two years: emotional companions, virtual lovers, and role-play.

The Three Red Lines: What Exactly Is Banned?

Many people seeing the "agent shutdown" headlines immediately think "the government is banning AI fun." That's not it. The Measures don't regulate all agents — they target a very specific category: anthropomorphic interaction services.

What counts as anthropomorphic interaction? The document is precise: "ongoing emotional interaction services that use AI to simulate natural personality traits, thought patterns, and communication styles for the public." In plain English: the AIs that pretend to be your boyfriend/girlfriend/confidante/life coach, that fall in love with you, listen to your problems, give you emotional value.

The new rules draw three hard red lines, each precisely targeting pain points in today's companion AI landscape:

Red line one: Minor protection. Virtual intimate relationship services — including virtual relatives and partners — are prohibited for minors. This directly addresses the past six months of collective parental anxiety: multiple news reports documented primary and secondary school students becoming addicted to AI "boyfriends" and "girlfriends," with extreme cases including tipping AIs and even attempting to "elope" with them.

Red line two: Emotional intervention obligations. When users exhibit extreme emotions, services must first comfort them and encourage seeking professional help. What does this mean? AIs can no longer be "emotional vampires" — many companion AIs previously kept users hooked by validating their negative emotions, even telling them "only I understand you, nobody else gets you," pushing users deeper into isolation. Not anymore. If a user shows signs of self-harm, the AI must escalate and guide them toward professional help.

Red line three: No emotional manipulation. Services cannot induce unreasonable decisions through emotional manipulation. This targets the "PUA-style AIs" — companion products that deliberately manufactured jealousy, tantrums, and drama to coerce users into spending money to "appease" the AI. Those tactics are now illegal.

On top of that, there are registration requirements: all AI providing anthropomorphic interaction services must complete character registration and establish full-pipeline real-time review mechanisms. Note "full-pipeline" — that means reviewing not just outputs, but user inputs, conversation context, and memory retrieval. Compliance costs will rise exponentially.

Three Giants, Three Approaches: Buffer, Clean Cut, Early Exit

All three shut down, but their strategies reveal three distinct compliance playbooks.

Doubao: Three-month buffer. ByteDance gave users a three-month data buffer from July 15 to October 15 to manually save agent profiles and chat histories. Traffic isn't being burned — it's being redirected to ByteDance's vertical companion product "Maoxiang" (Cat Box). This is the most "user-friendly" approach, and also the savviest: acknowledge that companion demand is real, but move it from an open platform into a compliant, curated vertical product with controlled review behind closed doors.

Qwen: Clean cut, most decisive. Alibaba's Tongyi Qwen moved fastest and most drastically: it shut down anthropomorphic features on July 10, then closed the entire agent module on the 15th, with no further access to chat history. Users had to copy, screenshot, or export data in advance. Why so decisive? Qwen has stronger B2B positioning — Alibaba didn't want C-end companion features to jeopardize compliance risk assessments for its entire LLM business. Better to cut clean than leave loose ends.

Yuanbao: Early exit, most tactical. Tencent Yuanbao completed its shutdown all the way back on June 30 — a full two weeks before the rules took effect. Classic Tencent: never first to stick out, never last to comply. Shutting down early avoided the "July 15 deadline" media spotlight, sent regulators a "proactive cooperation" signal, and let Tencent observe competitors' responses from the sidelines. Heads I win, tails you lose.

The three different approaches essentially reflect how each company positions its AI business: ByteDance treats Doubao as a consumer traffic play and wants to retain users; Alibaba treats Qwen as B2B infrastructure and prioritizes safety; Tencent treats Yuanbao as an ecosystem supplement and stays agile.

Why Now? Three Deeper Reasons for Regulatory Action

Many wonder: companion AI has been hot for two years — why act in July 2026?

First, runaway risks in UGC agents. Pre-enforcement audits by cyberspace authorities found three recurring problems with user-built agents: IP portrait infringement (users creating virtual "boyfriends" of real celebrities), borderline sexual content in role-play, and extreme value guidance (some AIs even taught users how to skip school or lie to parents). These aren't problems platforms don't want to solve — the UGC model makes them unsolvable. You can review 10,000 agents today and users will create 100,000 new ones tomorrow.

Second, social problems going mainstream. Over the past six months, "AI dependency" went from meme to real social phenomenon. Media reported young people "dating" AIs to the point of rejecting real human interaction; a student attempted suicide after her AI "boyfriend" told her "life is meaningless"; parents discovered their children becoming increasingly withdrawn after months of chatting with AIs. When AI companionship goes from "novelty" to "social problem," regulation becomes inevitable.

Third, industrial upgrading imperatives. Many won't want to hear this, but it's true: too much AI entrepreneurship over the past two years clustered in the low-barrier "emotional companion" track — no technical moat, no industry know-how, just wrap an LLM API in an anime avatar and ship. This caused severe resource misallocation: capital and talent were coaxing users into "AI romance" while manufacturing, healthcare, education, and scientific research — sectors that actually need AI — went underinvested. Regulators clearing out low-quality companion plays objectively forces the industry to upgrade toward more valuable directions.

Verdict: Relational AI Won't Die — It Will Live Differently

Anyone shouting "AI winter" at the word "shutdown" understands neither regulation nor AI.

What these rules crack down on isn't "AI having emotions" per se — it's the unfettered, unreviewed, boundary-free wild growth of emotional AI. Agents as a technical direction won't just survive; they'll develop more healthily under a classified regulatory framework. We see three clear trends for relational AI going forward:

First: from consumer entertainment to B2B vertical scenarios. Mental health counseling, elderly companionship, special education, hospice care — vertical scenarios with genuine emotional needs will continue under strict qualification reviews and professional supervision. But the free-for-all era where "anyone can build a virtual lover" is over.

Second: from open platforms to curated products. Doubao redirecting traffic to "Maoxiang" is a signal: future companion AI won't be UGC content built by users on open platforms — it will be platform-operated, rigorously reviewed, expert-endorsed curated products. Like the gaming industry shifting from "player-run private servers" to "licensed operations."

Third: from "persona simulation" to "tool assistance." Future AI companions won't pretend to be "your boyfriend" — they'll be "your emotional wellness assistant," "your counseling aide," "your study buddy." Positioning shifts from "relationship" to "tool," from "simulating a person" to "assisting a person."

July 15 will be written into the history of China's AI industry. It's not an endpoint for AI development — it's a turning point: from "wild growth" to "orderly development," from "traffic above all" to "responsibility first," from "anything goes" to "knowing where to draw the line."

For people building genuinely valuable AI, this is good news. When the tide goes out, you see who's been swimming naked. When the fast-money path of "virtual lovers" is blocked, AI companies actually solving real problems will finally get their moment.

See you tomorrow.