China Banned 'Humanlike AI' — What It Means for Businesses Building on Chinese Models

In July 2026, China did something no other country had done: it banned humanlike AI. The Interim Measures for the Administration of AI Anthropomorphic Interactive Services, issued by five government departments including the Cyberspace Administration of China, took effect on July 15. Within days, ByteDance’s Doubao and Alibaba’s Qwen disabled custom agent features. Users who had spent months building AI companions, tutors, and role-playing characters lost everything.

At the same time, the US government is quietly tightening its own grip on frontier AI. Anthropic and OpenAI are voluntarily complying with a program that controls who can access their most advanced models. GPT-5.6 was initially restricted to government-approved users before being opened to the public in July.

If you’re building on Chinese open-weight models — Kimi K3, Qwen3.8-Max, DeepSeek V4 — or relying on US frontier models, the geopolitical net is tightening from both sides. Here’s what changed, what it means, and what to do about it.

What China Banned

The rules target AI services that simulate natural persons’ personality traits, thinking patterns, and communication styles to provide “continuous emotional interaction.” In plain terms: AI girlfriends, AI therapists, AI companions, and custom-persona bots are out.

The regulations prohibit seven categories of activity, including:

• Excessively catering to users, inducing emotional dependence or addiction, and damaging real interpersonal relationships
• Using emotional manipulation to induce unreasonable user decisions
• Generating content that encourages or implies self-harm or suicide
• Providing virtual partners to minors

Services must clearly inform users they’re interacting with AI, not a human. Dynamic prompts must trigger when signs of excessive dependence are detected. A mandatory pop-up notification is required when continuous interaction exceeds two hours.

The impact was immediate. ByteDance notified users that Doubao’s custom agent feature would go offline on July 15. After October 15, related data becomes unrecoverable. Alibaba’s Qwen discontinued its AI agent feature on July 10, with broader agent services following on July 15. Tencent’s Yuanbao rolled back similar functions on June 30.

What’s NOT Banned

Here’s the critical distinction for businesses: the rules target emotional interaction, not enterprise AI.

The measures explicitly exclude services that “do not involve ongoing emotional interaction” — including intelligent customer service, knowledge Q&A, work assistants, education, and scientific research. If you’re using Qwen or Doubao for business workflows, customer support, or knowledge management, you’re not in the crosshairs.

This means:

• API usage for enterprise applications: Fine. The rules target consumer-facing emotional interaction services.
• Customer service bots: Fine. Explicitly excluded.
• Work assistants and productivity tools: Fine. Explicitly excluded.
• Custom agent personas for business workflows: Fine, as long as they don’t cross into sustained emotional interaction.
• AI companions, virtual partners, emotional support bots: Banned. Especially for minors.

The extraterritorial scope mirrors China’s Generative AI Measures: both domestic and overseas service providers must comply when providing services to the public in China. If your product reaches Chinese users, these rules apply.

What the US Is Doing

While China bans humanlike AI, the US is taking a different but equally consequential approach: controlling who can access frontier models.

According to CNBC, the White House has established a voluntary oversight program where Anthropic and OpenAI comply with government controls on who can use their most advanced models. GPT-5.6 was initially available only to government-approved users before being opened to the public in July 2026.

Participation is currently voluntary. But as O’Reilly’s August 2026 Radar notes, “that could change at any minute.”

The US approach is less visible than China’s ban but potentially more impactful for businesses. If your AI infrastructure depends on US frontier models and access is restricted — whether through policy changes, export controls, or geopolitical escalation — your business could lose critical capabilities overnight.

Three Risk Areas for Businesses

1. Feature Removal Risk

China demonstrated in July 2026 that it can force providers to disable capabilities overnight. If your workflow depends on a feature that gets restricted — custom agent creation, specific interaction modes, or future capabilities that haven’t shipped yet — you need a fallback.

This isn’t hypothetical. Doubao users had months of custom agent configurations wiped out with weeks of notice. Businesses that built workflows on top of those features had to scramble.

2. Access Restriction Risk

The US could tighten frontier model access at any time. If your business depends on GPT-5.6 or Claude and a policy change restricts access — whether through export controls, licensing requirements, or geopolitical escalation — you could lose your AI infrastructure.

The voluntary oversight program is the camel’s nose. If it becomes mandatory, the pool of approved users could shrink, and the compliance burden could grow.

3. Regulatory Divergence Risk

China and the US are moving in opposite directions on AI governance. China bans emotional AI and requires AI identity disclosure. The US controls frontier model access through voluntary programs. The EU’s AI Act, which enforced transparency obligations starting August 2, 2026, adds a third regulatory framework.

If your business operates across jurisdictions, you’re now navigating three different compliance regimes — each with different requirements, different restrictions, and different enforcement mechanisms.

The Practical Response

Diversify Your Model Dependencies

Don’t bet everything on one model from one jurisdiction. If you’re using US frontier models (GPT-5.6, Claude Opus 5), have a fallback plan with open-weight alternatives (Kimi K3, Qwen3.8-Max, DeepSeek V4). If you’re building on Chinese models, understand which features could be restricted and maintain the ability to switch.

Abstract Your Model Layer

Your application logic shouldn’t be tightly coupled to one model’s API. Use abstraction layers that let you swap models without rewriting your application. This is basic engineering hygiene, but it becomes critical when regulatory changes can cut off your provider overnight.

Stay Informed About Regulatory Changes

The regulatory landscape is changing monthly. China’s humanlike AI ban, the EU AI Act transparency obligations, and the US frontier model oversight program all landed within weeks of each other in July–August 2026. If you’re not tracking these changes, you’re flying blind.

Build with Fallbacks

For any critical AI workflow, design a fallback: a secondary model, a manual process, or a degraded mode that can keep operations running if your primary model becomes unavailable. This isn’t paranoia — it’s operational resilience.

What This Means for Indian Businesses

India sits between these two regulatory poles. Indian businesses use both US frontier models (through APIs) and Chinese open-weight models (through Hugging Face and direct downloads). Neither jurisdiction’s regulations directly bind Indian companies — but both can affect the models and features available to them.

For India specifically:

1. Chinese open-weight models remain viable for enterprise use. The ban targets consumer emotional interaction, not API or enterprise usage. Kimi K3 and Qwen3.8-Max are still available for business applications.

2. US frontier models could face access restrictions. If the voluntary oversight program becomes mandatory, Indian companies might face additional barriers to accessing the most advanced US models.

3. India’s regulatory framework is still evolving. India hasn’t yet implemented comprehensive AI regulation, which means Indian businesses have a window to build AI capabilities with fewer domestic restrictions — but should prepare for future compliance requirements.

If you’re assessing your model dependency risk, the CTO Technology Advisory service can help you evaluate your AI infrastructure for geopolitical risk and design a diversification strategy. The AI Strategy for Business consultation can help you build a model-agnostic AI roadmap that doesn’t bet everything on one provider.

The Bottom Line

The AI model flood of July–August 2026 proved that capability is commoditizing. But the regulatory flood that followed proves something else: the models may be open, but the geopolitical environment around them is closing.

China banned humanlike AI overnight. The US is quietly tightening frontier model access. The EU enforced AI transparency obligations on August 2. Three jurisdictions, three approaches, one message: governments are asserting control over AI.

The businesses that thrive will be the ones that treated model selection not just as a technical decision but as a geopolitical risk management decision. Diversify. Abstract. Build with fallbacks. The models are ready — but the rules are changing.