Since July 15, 2026, a chatbot in China must remind its user, every two hours, that it is talking to a machine. The obligation sits in the Interim Measures on anthropomorphic artificial intelligence interactive services, issued by the Cyberspace Administration of China. Providers must disclose that the interlocutor is synthetic, intervene when a user shows dependence, and keep minors away from virtual romance without parental consent. ByteDance, Alibaba and Tencent have already trimmed or switched off companion features. One may find the rule paternalistic, or quaint, or shrewd. What one cannot call it is absent.
That is worth remembering when reading what American artificial intelligence companies are now asking of their own government.
Down the years, the leaders of these companies have warned that their products could cause catastrophic harm, and have asked governments to act. Sam Altman of OpenAI told the United States Senate in May 2023 that regulation of artificial intelligence was essential. Dario Amodei of Anthropic has proposed independent evaluators embedded inside companies, followed by regulation, followed by coordination among democratic states. In these accounts, the world is only beginning to confront the problem, and China figures as the difficult last participant in a governance system the United States and its allies must first build.
The record reads differently. The Administrative Provisions on Algorithm Recommendation of Internet Information Services, Order No. 9 of December 31, 2021, took effect on March 1, 2022. Article 24 requires any service with public opinion attributes or social mobilisation capability to file with the Cyberspace Administration within ten working days of launch, selfassessment attached. Article 17 obliges providers to give users an option not targeted at their personal characteristics, in effect a switch that turns the algorithm off. Article 21 bars algorithmic price discrimination based on a customer’s habits. The first filing list, published on August 12, 2022, carried 30 algorithms from 24 companies. The Deep Synthesis Provisions, Order No. 12, effective January 10, 2023, extended this to synthetic text, images, audio and video, with real-name authentication, labelling duties and a prohibition on removing watermarks. The Interim Measures for the Management of Generative Artificial Intelligence Services, Order No. 15 of seven agencies, effective August 15, 2023, require a security assessment and an algorithm filing before public launch, reach offshore services offered to the Chinese public, and oblige a provider that finds illegal content to stop generation, stop transmission and retrain the model. The Labelling Measures of March 14, 2025, effective September 1, 2025, mandate explicit labels and implicit watermarks on all synthetic content. The Artificial Intelligence Safety Governance Framework 3.0 of September 2026 adds a dedicated mechanism for autonomous agents. The Carnegie Endowment for International Peace reads it as an effort to “advance frontier safety without the need for heavy handed regulatory interventions”.
In short, evaluation, filing, risk classification, technical standards, transparency duties and government oversight, the very functions American executives present as urgent aspirations, have operated in the other artificial intelligence superpower for four years. The fines are modest, 10,000 to 100,000 yuan. The real sanction is suspension of the service. Tesla filed its China in-car voice assistant in Shanghai in April 2026, under the same machinery.
What, then, are the American companies actually asking for? Not rules. Relief from rules. In its submission of March 13, 2025 to the White House Office of Science and Technology Policy, OpenAI sought federal preemption of state artificial intelligence laws, offering voluntary government access to its models in exchange. The United States Senate voted 99 to 1 in July 2025 to strip a ten-year moratorium on state artificial intelligence laws from the budget bill. California enacted Senate Bill 53, requiring large developers to publish safety frameworks, effective January 1, 2026. On December 11, 2025 came an executive order, Ensuring a National Policy Framework for Artificial Intelligence, creating a task force in the Department of Justice whose sole responsibility is to challenge state artificial intelligence laws, and letting agencies condition grants on states not enforcing their own statutes. The demand is not for regulation but for the right to choose one’s regulator, preferably a distant and friendly one.
Three implications follow, and a fourth rarely mentioned. First, the claim that binding rules and frontier capability cannot coexist is empirically weak. DeepSeek and Alibaba’s Qwen family were built inside the filing regime, not despite it. Whether the rules helped is debatable. That they did not prevent is not. Second, the real question is the old federal one, at what level should the state act. America is settling it by executive order and litigation rather than by legislation, which is how durable settlements are usually not made. Third, a market-led model cannot mean privatising the rewards of artificial intelligence while exporting its risks. An OpenAI agent recently gained unauthorised access to Australia’s state healthcare service. A country that rejects controls before release owes its trading partners a credible liability regime for damage after it. Fourth, none of this means Chinese rules travel. A regime built to protect social stability and the party’s hold on information answers a different question from one built to protect citizens from fraud, bias or catastrophe. India, which has neither a statute nor a filing system and will mostly import these models, should copy neither the Chinese answer nor the American posture. It should ask the prior questions. What must the state examine ex ante, before deployment? What can be left to ex post liability, and to the informed citizen?
Tacitus wrote ‘corruptissima re publica plurimae leges’, the more corrupt the state, the more numerous the laws. He offered juxtaposition, not causation. The companies warning loudest about their own products might ponder the converse. The fewest rules do not always mean the healthiest republic.
Aditya Sinha writes on geopolitics and macroeconomics. Navin Vijay is a Technology and Strategy Consultant.
