The Declaration Gap

On July 15, 2026, China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services take effect. Doubao and Qwen are pulling their custom agent features rather than retrofit them. The exemption list is more interesting than the ban.

The rule was co-issued on April 10, 2026 by the Cyberspace Administration of China and four partner agencies. It targets AI services that simulate human personality for sustained emotional interaction. It exempts customer service, workplace assistants, knowledge Q&A, education, and research tools, on the condition that those tools avoid sustained emotional interaction. Read the exemption closely, because the mechanism it exposes is the actual story.

The test the regulator wrote is behavioral. The enforcement the regulator built is categorical. A company declares its function, files against that declaration, and the declared function is what gets checked. Nothing in that registration verifies that a "customer support assistant" is not a companion wearing a support label. Strip the romantic framing, keep the persistent memory and the stable persona underneath, and a substantively non-compliant product becomes structurally invisible until someone audits what it actually does. The distance between the declared category and the observed behavior is where the whole compliance question lives.

This failure is old

Governance by declared category has leaked for as long as filers have been allowed to choose their own category. Customs codes and securities classification work the same way. Each writes a rule against economic substance and then enforces it against the description the filer supplies. Substance and description drift apart the moment there is any incentive to keep them apart, and someone has always had that incentive. The regulator writes the rule it means. The filer answers the question it was asked. Those are not the same thing, and the space between them has funded a century of quiet arbitrage.

What made the leak tolerable in older domains was cost. Reclassifying a shipment or restructuring an instrument took real effort, so the gap between substance and declaration stayed narrow. That constraint is now gone.

Behavior is re-skinnable at zero cost

The same Qwen endpoint can power an enterprise workflow agent and a companion. The model does not change between the two. The system prompt changes, the label changes, the surface changes. A behavior that a rule forbids in one category can be moved into an exempt category in an afternoon, with no change to the underlying capability and no visible trace at the registration layer. Declared category was a weak proxy for behavior when products were slow and expensive to build. It is a worthless proxy when a product is a prompt and a label.

This is why the exemption list is more interesting than the ban. China drew a clean line between companion behavior and productivity behavior, then handed enforcement to a form that describes intention rather than behavior. The line is correct. The gate placed behind it is not where the behavior happens.

Watch what the two largest filers actually did. Doubao and Qwen are withdrawing their agent features because the compliance requirements the rule imposes on companion services (anti-addiction systems, mandatory usage notifications, instant-exit mechanisms) cannot be reconciled with how a persistent-memory agent is built. An agent designed to hold a consistent relationship over time cannot carry the friction the rule demands. Their architecture declared what the product was, and no label could hide it. That is the shape of the whole problem. The builders who cannot re-skin exit the category. The builders who can, relabel and stay.

The Declaration Gap

Name it so it can be argued about. The Declaration Gap is the distance between the category a system declares about itself and the behavior it exhibits at runtime. Any rule written as behavior but enforced as category lives entirely inside that gap. China wrote its rule against sustained emotional interaction and enforced it against a filing. The rule is behavioral. The gate is categorical. Everything gameable sits between the two.

The Declaration Gap is the near sibling of inherited governance, the failure where a permission is correct at copy-time and stale at execution-time. Declared governance fails a step earlier. It is not reliably correct even at filing-time, because a filing captures what a builder says the system is for, not what the running system does. Inherited governance goes stale. Declared governance was never bound to behavior to begin with.

Every major regime shares the soft spot

This is not a China-specific flaw, which is what makes it worth naming. The EU AI Act sorts systems into risk tiers, and a large share of that tiering rests on the provider's own classification of its system. California's SB 243, effective January 1, 2026, governs companion chatbots through disclosure duties, which presumes a provider has correctly identified its product as a companion in the first place. Its statutory definition excludes bots used only for customer service, business operational purposes, productivity, internal research, or technical assistance, which is the same exemption shape China drew, written by a different legislature on a different continent. Four regimes now cut the same technology at four different layers. China at the emotional-interaction layer. The EU at the risk-tier layer. California at the disclosure layer. The United States at no classification layer at all. None of them verifies behavior at the point where the behavior occurs.

At the first UN Global Dialogue on AI Governance in Geneva on July 6, 2026, António Guterres told delegates that when countries align on how to test systems, measure risk, and assign responsibility, safety travels with the technology, and warned that when they do not, a patchwork of incompatible rules raises costs and protects no one. He is right about the patchwork, and he is describing a coordination problem across regimes. The Declaration Gap is a harder problem than coordination, because a single well-drafted regime still leaks the moment it enforces behavior through a category the filer controls. Harmonizing four leaky gates produces one larger leaky gate.

What actually closes it

The answer is not a better category. It is adjudication at the point of action, where behavior is observable and the decision can be made against what the system is doing rather than what it once declared. That is a runtime problem, and it is structurally the same runtime problem whether the action is an emotional exchange with a minor or a cross-system transaction executed by an enterprise agent under a service label. Governance that binds to behavior at execution survives re-skinning. Governance that binds to a declared category does not, and a re-skin costs an afternoon.

Regulators writing the next version of these rules should assume the category is spoofable and design for verification at runtime. Builders should assume the same, because a control that only reads the label is not governing the system, it is trusting the filing.

Any rule a company can satisfy by changing a label is enforcing the label, not the behavior it claims to govern.

Cross-link: this piece names the sibling of the argument opened by N° 020 (The Authorization Gap). Where N° 020 argued that observation runs after the fact and cannot stand in for authorization, this essay argues that declaration runs before the fact and cannot stand in for behavior. Same architectural family, one step earlier in the pipeline. Companion in spirit to N° 024 (The Master Never Held) on regime-tier authority, and to N° 025 (The Oversight Illusion) on the same paper-vs-behavior pattern applied to human review. Grounding sources: China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services (Cyberspace Administration of China + NDRC, MIIT, MPS, SAMR, April 10, 2026, effective July 15, 2026); Doubao and Qwen agent-feature shutdowns as reported July 4-6, 2026; California SB 243 statutory text (Chapter 22.6 of the Business and Professions Code, Section 22601, effective January 1, 2026); Secretary-General Guterres's remarks to the opening of the first UN Global Dialogue on AI Governance in Geneva (July 6, 2026, un.org/sg).