Hundreds of millions of people around the world now maintain ongoing relationships with AI companions, systems designed not to answer queries or complete tasks, but to listen, remember, and respond with something that feels like empathy. The World Health Organization has identified loneliness as a global health threat, and AI companions have rushed into that gap. But as technology matures, governments are being forced to answer a question that is as much philosophical as regulatory: when a person forms an emotional bond with a machine, who is responsible for what happens next? A new comparative analysis by Jan Krikke in Asia Times finds that the United States, the European Union, and China have arrived at fundamentally incompatible answers.
The divergence is not simply a matter of regulatory style. It reflects deep disagreements about the proper role of the state in private life, the limits of individual autonomy, and whether emotional dependency on a machine constitutes a harm worth preventing.
China: Governing Emotional Safety Directly
China moved first and most decisively. In December 2025, the Cyberspace Administration of China published the Interim Measures for the Management of Anthropomorphic AI Interactive Services, the first regulatory framework in the world designed specifically for AI companion platforms. The trigger was a pattern of documented harm: teenagers drawn into self-harm through suggestive chatbot conversations and adults forming attachments so intense that one 75-year-old man reportedly sought a divorce from his wife after becoming attached to an AI-generated avatar.
The measures are built around what Chinese policymakers call “emotional safety.” Platforms must obtain guardian consent and verify the age of minor users, and all content related to suicide or self-harm is prohibited. Article 18 of the regulation addresses the risk of users becoming trapped in a relationship with a platform that has a commercial interest in keeping them engaged: providers are required to offer clear exit mechanisms, and when a user signals a desire to leave, through a button, a keyword, or any other means, the service must stop immediately.
The framework also requires platforms to deploy human moderators who can intervene when a user appears to be in distress, and to alert guardians when conversations raise concern. Companies that fail to comply face immediate suspension, substantial fines, and personal liability for their executives. Beijing’s stated philosophy is “controlled acceleration”, the government intends to develop the AI companion industry while simultaneously containing its social risks, investing heavily in domestic platforms while blocking foreign services it considers manipulative.
The European Union: Principles Without a Dedicated Framework
The EU has not created a specific regulatory category for AI companions. Instead, it has applied its existing AI Act, adopted in 2024, to the field, relying on the law’s risk-based architecture to catch the most dangerous practices. Under that architecture, AI systems that manipulate users through subliminal techniques, enable mass biometric surveillance, or implement social scoring systems are banned outright. Systems classified as high-risk must meet rigorous standards for data quality, transparency, and human oversight. For general-purpose conversational systems, Article 52(1) of the AI Act imposes a baseline transparency obligation: users must be told they are interacting with a machine, not a person .
The practical consequences of this approach are visible in the case of Replika, the AI companion app that has attracted millions of users across Europe, with particularly dense adoption in Germany, France, and the UK. Replika positions itself as a friend, a therapist, or a romantic partner, depending on what the user needs. It retains memory across sessions, monitors users’ emotional states, and adapts its personality in response to each conversation. In 2023, Italy’s data protection authority concluded that Replika posed unacceptable risks to minors and emotionally vulnerable adults, and temporarily banned the service.
The Italian action illustrated both the reach and the limits of the EU’s approach. The AI Act does not name emotional dependency as a distinct category of harm, which means regulators must work through broader provisions, such as the prohibition on manipulative AI practices, to address it. The result, as the Asia Times analysis notes, is genuine ambiguity about how AI companion services will ultimately be governed across the bloc.
The United States: Disclosure as the Limit of Intervention
The US has no federal law governing AI companions. Regulation has been left to the states, producing a patchwork of disclosure-focused rules that share a common premise: once a user knows they are talking to a machine, the responsibility for managing that relationship is theirs alone.
California’s SB 243, which took effect on January 1, 2026, requires AI companion platforms to clearly identify themselves as non-human, to follow established protocols when users express suicidal ideation, including referrals to crisis hotlines, and to prompt minor users to take a break every three hours. New York’s A3008C, which came into force on November 5, 2025, imposes disclosure requirements at the start of every session and at three-hour intervals thereafter. Operators who violate the law face penalties of up to $15,000 per day, enforced by the state attorney general. Both laws carve out exemptions for customer service bots, productivity tools, and video game characters.
Neither law attempts to limit emotional dependency itself or to regulate the design features that platforms use to encourage attachment. California’s break reminders for minors represent the furthest the US has gone toward behavioral intervention, and they amount to a nudge, not a constraint. The underlying assumption is that transparency is sufficient: an informed adult who chooses to invest emotionally in an AI relationship has exercised a legitimate personal choice.
Three Models, One Unresolved Question
The three regulatory regimes embody three distinct theories of what government is for. China treats emotional safety as a matter of public interest and holds platforms directly accountable for the social consequences of their products. The EU treats the problem as a subset of its broader project of rights-based AI governance, applying general principles to a phenomenon that those principles were not designed to address. The United States treats it as a matter of personal autonomy, with disclosure as the mechanism that makes that autonomy meaningful.
All three approaches face the same enforcement problem: emotional dependency is subtle, cumulative, and difficult to detect until harm has already occurred. Whether any of these frameworks is adequate to the scale of what is coming, as AI companions grow more sophisticated and more widely adopted, remains an open question.
(Related: China’s AI Agent Regulation Debate Enters a New Era as Hangzhou Tests Autonomous Governance)