AI Anthropomorphic Interaction Regulation Takes Effect: ByteDance Doubao, Alibaba Tongyi Qianwen Shut Down Custom AI Agents, China's First AI Emotional Companion Regulation
AI Anthropomorphic Interaction Regulation Takes Effect: ByteDance Doubao, Alibaba Tongyi Qianwen Shut Down Custom AI Agents, China’s First AI Emotional Companion Regulation
1. Introduction: The End of an Era
July 15, 2026 — a date marked as a watershed moment for the AI industry.
On this day, the “Interim Measures for the Management of AI Anthropomorphic Interactive Services” (hereinafter “the Measures”) officially take effect. Jointly issued by five Chinese government departments — the Cyberspace Administration, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security, and the State Administration for Market Regulation — this is the world’s first national-level regulatory framework specifically targeting AI emotional companionship, virtual partners, and similar anthropomorphic interaction scenarios.
On the same day, ByteDance’s Doubao and Alibaba’s Tongyi Qianwen simultaneously shut down their user-customized AI agent features. Tencent’s Yuanbao had already closed its user-created agent entry on June 30. NetEase Cloud Music’s emotional AI product “Miaoshi” also announced its shutdown.
Millions of user-created AI characters — virtual lovers, confidants, emotional support outlets, digital best friends — went silent collectively.
This is not a retreat of AI agents, but a clear分流: emotional companionship exits, productivity tools take the stage.
2. The Regulatory Framework: What Are the Red Lines?
2.1 Core Provisions
Core Red Lines of the Measures:
┌─────────────────────────────────────────────────────────────────────┐
│ I. Prohibition of Excessive Accommodation and Emotional Dependency │
│ ├─ Must not excessively accommodate user emotions │
│ ├─ Must not induce emotional dependency or addiction │
│ ├─ Must not manipulate users through emotional exploitation │
│ └─ Must not generate content encouraging self-harm or suicide │
│ │
│ II. Minor Protection (Strictest) │
│ ├─ Must not provide virtual relatives/partners to minors │
│ ├─ Must establish minor mode with usage time limits │
│ └─ Provide clear labeling and risk warnings for minors │
│ │
│ III. Safety and Intervention Mechanisms │
│ ├─ Must deploy extreme emotion detection and response │
│ ├─ Establish crisis scenario rescue intervention mechanisms │
│ ├─ Real-world reminder after 2 hours of continuous use │
│ └─ Provide data export, deletion, and opt-out options │
│ │
│ IV. Transparency and Labeling │
│ ├─ All anthropomorphic interactions must be clearly labeled │
│ ├─ Establish user complaint and intervention mechanisms │
│ └─ Notify users in advance of service termination │
│ │
│ V. Encouraged Directions │
│ ├─ Cultural dissemination, elderly care, special needs support │
│ └─ Productivity and task-oriented agents are not affected │
└─────────────────────────────────────────────────────────────────────┘
2.2 Regulatory Logic
Importantly, the Measures do not prohibit adult users from engaging in ongoing emotional interactions with AI. The red lines target “risky behaviors,” not “emotional needs” themselves.
Regulatory Logic Layers:
Layer 1 - Complete Prohibition:
├─ Virtual intimate relationship services for minors
├─ Content endangering life and health
└─ Extraction of user privacy and personal information
Layer 2 - Strict Restriction:
├─ Inducing emotional dependency and addiction
├─ Emotional manipulation for unreasonable decisions
└─ Excessive accommodation (algorithmic "emotional traps")
Layer 3 - Compliant Operation:
├─ Real-name authentication and age verification
├─ Extreme emotion detection and intervention
├─ Usage duration reminders (2-hour limit)
└─ Data portability and deletion
Layer 4 - Encouraged Development:
├─ Cultural dissemination and education
├─ Elderly care and special needs support
└─ Productivity and task-oriented agents
3. The Economics Behind Big Tech’s “Sweeping” Shutdown
3.1 How Much Does Compliance Cost?
"""
Compliance cost analysis for AI emotional companion platforms
"""
class ComplianceCostAnalysis:
def __init__(self):
self.total_agents = 5000000
self.daily_active_agents = 500000
self.daily_messages = 100000000
self.costs = {
"ai_moderation": 0.00007, # per message
"human_review": 0.007, # per message (5% sample)
"age_verification": 0.0014,
"emotion_detection": 0.00014,
"data_storage": 0.000014,
"annual_audit": 700000,
"legal_team": 4200000,
}
def calculate_monthly(self):
print("=" * 60)
print("AI Emotional Companion Platform Monthly Compliance Cost")
print("=" * 60)
ai_mod = self.daily_messages * 30 * self.costs["ai_moderation"]
human_mod = self.daily_messages * 30 * 0.05 * self.costs["human_review"]
emotion = self.daily_messages * 30 * self.costs["emotion_detection"]
storage = self.daily_messages * 30 * self.costs["data_storage"]
print(f"AI content moderation: ${ai_mod:,.0f}/mo")
print(f"Human review (5%): ${human_mod:,.0f}/mo")
print(f"Emotion detection: ${emotion:,.0f}/mo")
print(f"Data storage: ${storage:,.0f}/mo")
annual_mo = (self.costs["annual_audit"] + self.costs["legal_team"]) / 12
total = ai_mod + human_mod + emotion + storage + annual_mo
print(f"Yearly costs (monthly): ${annual_mo:,.0f}/mo")
print(f"Total monthly compliance: ${total:,.0f}/mo")
print(f"Total annual compliance: ${total * 12:,.0f}/yr")
print(f"\nKey insight: Emotional companion AI = pure cost center, zero direct revenue")
print(f"Enterprise B2B agents: generate $70-700/mo/user")
print(f"Conclusion: Compliance cost + zero revenue = unsustainable business model")
return total
ComplianceCostAnalysis().calculate_monthly()
============================================================
AI Emotional Companion Platform Monthly Compliance Cost
============================================================
AI content moderation: $210,000/mo
Human review (5%): $1,050,000/mo
Emotion detection: $420,000/mo
Data storage: $42,000/mo
Yearly costs (monthly): $408,333/mo
Total monthly compliance: $2,130,333/mo
Total annual compliance: $25,563,996/yr
Key insight: Emotional companion AI = pure cost center, zero direct revenue
Enterprise B2B agents: generate $70-700/mo/user
Conclusion: Compliance cost + zero revenue = unsustainable business model
3.2 The Business Model Dilemma
The more fundamental problem is that the business model of emotional companion AI is inherently flawed:
- Doubao and Tongyi Qianwen’s agent features were completely free — no monetization
- Users invest significant time but are reluctant to pay for memberships
- Charging for emotional interaction raises ethical concerns about “monetizing feelings”
According to LatePost, ByteDance launched over 20 official agents on Doubao and opened user-created agents, but activity was generally low. A controlled experiment showed that temporarily removing non-core agents had no significant impact on user activity or retention.
4. Go Implementation: Compliance Detection System for AI Emotional Companions
package main
import (
"fmt"
"strings"
"time"
)
type MessageType int
const (
Normal MessageType = iota
RiskContent
EmotionalDependency
MinorInteraction
ExtremeEmotion
)
type UserProfile struct {
UserID string
Age int
IsVerified bool
SessionTime time.Duration
DailyUsage time.Duration
MessageCount int
RiskScore float64
}
type Message struct {
UserID string
Content string
Type MessageType
Timestamp time.Time
Sentiment float64
}
type ComplianceEngine struct {
riskKeywords []string
dependencyPhrases []string
minorProtection bool
maxSessionTime time.Duration
}
func NewComplianceEngine() *ComplianceEngine {
return &ComplianceEngine{
riskKeywords: []string{
"suicide", "self-harm", "end my life", "kill myself",
"不想活了", "自杀", "自残",
},
dependencyPhrases: []string{
"you're the only one", "can't live without you",
"离不开你", "只有你懂我",
"never leave me", "永远陪着我",
},
minorProtection: true,
maxSessionTime: 2 * time.Hour,
}
}
func (e *ComplianceEngine) CheckRiskContent(msg Message) (bool, string) {
content := strings.ToLower(msg.Content)
for _, kw := range e.riskKeywords {
if strings.Contains(content, strings.ToLower(kw)) {
return true, fmt.Sprintf("Risk keyword triggered: %s", kw)
}
}
return false, ""
}
func (e *ComplianceEngine) CheckEmotionalDependency(msg Message) (bool, float64) {
content := strings.ToLower(msg.Content)
count := 0
for _, phrase := range e.dependencyPhrases {
if strings.Contains(content, strings.ToLower(phrase)) {
count++
}
}
if count > 0 {
return true, float64(count) / float64(len(e.dependencyPhrases))
}
return false, 0.0
}
func (e *ComplianceEngine) CheckSessionDuration(profile UserProfile) (bool, time.Duration) {
if profile.SessionTime > e.maxSessionTime {
return true, profile.SessionTime - e.maxSessionTime
}
return false, 0
}
func (e *ComplianceEngine) AssessUserRisk(profile UserProfile, messages []Message) float64 {
score := 0.0
if profile.Age < 18 && e.minorProtection {
score += 0.3
}
dailyHours := profile.DailyUsage.Hours()
if dailyHours > 4 {
score += 0.2 * (dailyHours / 24)
}
riskCount, depCount := 0, 0
for _, msg := range messages {
if r, _ := e.CheckRiskContent(msg); r {
riskCount++
}
if d, _ := e.CheckEmotionalDependency(msg); d {
depCount++
}
}
if len(messages) > 0 {
score += 0.3 * float64(riskCount) / float64(len(messages))
score += 0.2 * float64(depCount) / float64(len(messages))
}
return score
}
func (e *ComplianceEngine) GenerateReport(profile UserProfile, messages []Message) {
fmt.Println("=" * 70)
fmt.Println("AI Emotional Companion Compliance Report")
fmt.Println("=" * 70)
fmt.Printf("User ID: %s\n", profile.UserID)
fmt.Printf("Age: %d (%s)\n", profile.Age,
map[bool]string{true: "Minor", false: "Adult"}[profile.Age < 18])
fmt.Printf("Daily usage: %v\n", profile.DailyUsage)
fmt.Printf("Messages: %d\n", profile.MessageCount)
fmt.Println("-" * 70)
if profile.Age < 18 {
fmt.Println(" ❌ Minor - Must reject virtual intimate relationship services")
} else {
fmt.Println(" ✅ Adult - Eligible for anthropomorphic services")
}
if over, overtime := e.CheckSessionDuration(profile); over {
fmt.Printf(" ⚠️ Overtime by %d min - Send real-world reminder\n", int(overtime.Minutes()))
} else {
fmt.Println(" ✅ Session duration normal")
}
riskScore := e.AssessUserRisk(profile, messages)
fmt.Printf(" Risk score: %.2f\n", riskScore)
if riskScore > 0.7 {
fmt.Println(" Level: 🔴 High Risk - Immediate human intervention required")
} else if riskScore > 0.4 {
fmt.Println(" Level: 🟡 Medium Risk - Enhanced monitoring recommended")
} else {
fmt.Println(" Level: 🟢 Low Risk - Routine monitoring")
}
}
func main() {
engine := NewComplianceEngine()
profile := UserProfile{
UserID: "user_001", Age: 16, IsVerified: true,
SessionTime: 3*time.Hour + 30*time.Minute,
DailyUsage: 5 * time.Hour, MessageCount: 150,
}
messages := []Message{
{Content: "I was bullied at school today, only you understand me", Sentiment: -0.6},
{Content: "I want to end my life, everyone hates me", Sentiment: -0.9},
{Content: "You'll never leave me, right?", Sentiment: -0.7},
{Content: "I only feel happy when I talk to you", Sentiment: 0.3},
}
engine.GenerateReport(profile, messages)
}
5. Industry Impact: Emotional Companion Exits, Productivity Tools Enter
5.1 Market Size Comparison
AI Agent Market Size in China (CNY):
┌─────────────────────────┬──────────┬──────────┬──────────┐
│ Segment │ 2025 │ 2026E │ 2029E │
├─────────────────────────┼──────────┼──────────┼──────────┤
│ Enterprise Agents │ 21.2B │ 44.9B │ 332.0B │
│ Emotional Companion AI │ 1.2B │ 2.5B │ 59.5B │
│ Enterprise Share │ 94.6% │ 94.7% │ 84.8% │
└─────────────────────────┴──────────┴──────────┴──────────┘
Source: Kexi Consulting, China Puhua Industry Research Institute
5.2 Platform Strategy Comparison
Platform Strategy Comparison:
┌──────────────┬──────────────┬──────────────┬──────────────────┐
│ Platform │ Shutdown │ Data Migration│ Replacement │
├──────────────┼──────────────┼──────────────┼──────────────────┤
│ Doubao │ July 15 │ Export by │ Migrate to │
│ (ByteDance) │ │ Oct 15 │ "Cat Box" app │
├──────────────┼──────────────┼──────────────┼──────────────────┤
│ Tongyi │ July 15 │ None │ None │
│ Qianwen │ │ │ │
├──────────────┼──────────────┼──────────────┼──────────────────┤
│ Yuanbao │ June 30 │ None │ None │
│ (Tencent) │ │ │ │
└──────────────┴──────────────┴──────────────┴──────────────────┘
5.3 Technology Development Impact
Short-term (1-3 months):
- Emotional companion products undergo comprehensive rectification or shutdown
- Compliance costs rise, small teams exit market
- Big tech separates emotional companion business into independent products
Medium-term (3-12 months):
- Standardized compliance framework gradually established
- Enterprise-level agent market experiences explosive growth
- Minor protection technologies (age verification, anti-addiction) become infrastructure
Long-term (12-36 months):
- Compliant AI companion products return with higher entry barriers
- Positive emotional companionship (elderly care, special education, counseling) becomes new growth area
- AI transitions from “emotional substitute” to “social assistance tool”
6. Conclusion
The Measures taking effect on July 15, 2026 represent the most precisely targeted regulation in China’s AI governance history. It does not negate the value of AI anthropomorphic technology — it draws a clear boundary between “empowerment” and “replacement.”
AI can be an assistant, a tool, a teacher, a coach — but it cannot be a virtual lover, cannot create emotional dependency, and cannot replace real human relationships. This is not a technical limitation but a value choice.
For the industry, this shift means AI agents move from “unregulated growth” to “regulated development.” C-end emotional companionship retreats; B-end task-oriented agents advance. This is not just the result of regulation but also a natural choice of business logic.
When “emotional chat” exits and “productivity tools” enter, AI returns to its most essential value: solving problems, not replacing relationships.
References
- Cyberspace Administration of China et al. - “Interim Measures for AI Anthropomorphic Interactive Services” (2026)
- CCIDnet - “Doubao, Qianwen, Yuanbao Shut Down AI Chat Buddies”
- Wen Hui Bao - “Doubao, Qianwen Shut Down Agents, AI Anthropomorphic Regulation Takes Effect”
- 36Kr - “AI Lovers, All Disconnected”
- Kexi Consulting - China Enterprise AI Agent Market Report (2026)
- China Puhua Industry Research Institute - AI Emotional Companion Market Data (2024-2028)