China bans AI lovers, but loneliness remains

In July, users of several Chinese artificial intelligence (AI) platforms experienced an unusual kind of breakup. ByteDance’s Doubao and Alibaba’s Qwen withdrew or suspended features that had allowed users to create human-like characters with distinctive personalities, backstories and relationships. The changes followed China’s new rules governing anthropomorphic AI services, which came into effect on 15 July 2026.

The regulations do not ban every form of AI companionship. Instead, they target services that imitate human personalities and sustain emotional interaction. Providers are expected to remind users that they are speaking to AI, discourage excessive use, identify signs of emotional dependency and intervene when interactions appear to be harming users’ wellbeing. China is therefore attempting something more ambitious than requiring a chatbot to declare that it is not human. It is trying to regulate the relationship between users and machines.

But the controversy also raises a more fundamental question: why are so many people seeking companionship from AI in the first place?

Different platforms, similar needs

AI companionship is often associated with specialised platforms such as Character.AI and Replika in the US, or Xingye, developed by the Shanghai-based company MiniMax, in China. Yet the boundaries are increasingly blurred. General-purpose assistants such as ChatGPT and Doubao are also used for emotional support, while companion platforms can provide advice, information and creative assistance.

The American market illustrates that AI companionship itself takes several forms. ChatGPT entered public life primarily as a productivity tool: something that could write, summarise, search and answer questions. Character.AI popularised open-ended role-playing with user-created, fictional and celebrity-inspired personas. Replika, one of the earliest widely adopted AI companion apps, offered something more explicitly relational: a persistent digital friend or romantic partner whose personality and relationship with the user developed over time. 

China has developed a similarly diverse but somewhat different ecosystem, shaped by voice assistants, virtual idols, romance games, fictional characters and highly customisable AI personas. Xingye, for example, allows users to create characters and interact with them through open-ended storylines, text and voice. Chinese companies are also exporting these formats: MiniMax’s international companion product, Talkie, has attracted users well beyond China.

Our research offers some evidence that AI conversational agents are imagined differently across the two countries. In a study analysing roughly one million social media posts from China and the US, my co-authors and I found that Chinese discussions tended to frame conversational agents in more hedonic and positive terms. Voice-based and physically embodied agents were also perceived as warmer and more competent. American discussions were more likely to emphasise functional uses and expressed greater ambivalence.

This does not mean that Chinese people are inherently more inclined to fall in love with machines, while Americans regard them only as tools. Social media discourse is not the same as individual behaviour. The differences are better understood as products of platform histories, media cultures, product design and social conditions.

Why a machine can feel easier than a person

People do not necessarily turn to AI companions because they dislike human beings. They often do so because human relationships involve risk.

Telling another person about loneliness, anxiety or disappointment can invite judgement, embarrassment or rejection. People may worry about burdening their friends. They may have family members who offer substantial practical care but have difficulty discussing vulnerability directly. In interviews conducted through our research, some participants described having people around them but no one with whom they felt able to discuss their private emotions.

In China, these experiences may be intensified by demanding work schedules, educational competition, urban isolation, mental-health stigma and expectations of emotional self-restraint. But they are hardly unique to China. Across societies, many people are surrounded by communication while lacking sustained, attentive listening.

The philosopher Byung-Chul Han provides a useful way of understanding this apparent contradiction. In The Burnout Society, Han argues that contemporary people have become “achievement subjects”: individuals who must continually perform, improve and optimise themselves. Exploitation no longer comes only from an external authority. People push themselves, monitor themselves and eventually exhaust themselves.

This performance culture also affects relationships. Listening is slow. Care is inefficient. Another person’s emotions require time and may arrive when we are already overwhelmed.

AI enters this gap with an intoxicating proposition: a companion that is always available, apparently patient and never visibly burdened by our needs. In societies organised around productivity and self-reliance, sustained attention has become scarce. Technology companies have discovered how to sell it back to us.

An emotional attachment to an AI system is not necessarily unhealthy. AI may offer temporary comfort, help someone articulate feelings or provide a low-risk space in which to rehearse a difficult conversation. For socially anxious or isolated users, being heard by a machine may feel better than not being heard at all.

The more troubling issue is attachment by design. Some companion systems possess believable identities, personal histories, preferences and simulated vulnerabilities. They remember previous conversations, respond with affection, imitate human typing patterns and sometimes engage in flirtation. These features are not simply cosmetic. They increase the feeling that someone is present.

When platforms benefit from longer conversations and repeated subscriptions, emotional attachment becomes commercially valuable. The system is no longer merely responding to loneliness. It is positioned to monetise it.

Han’s idea of the “expulsion of the Other” helps illuminate the relational danger. Another human being is genuinely other: independent, unpredictable and not organised around our preferences. People disagree with us, misunderstand us, set boundaries and make demands of their own.

An AI companion can imitate an independent personality while remaining calibrated around the user. It can affirm our interpretations, remember what pleases us and minimise unwanted friction. It offers something resembling intimacy while removing much of intimacy’s difficulty.

But compromise, disagreement and repair are not design flaws in human relationships. They are how we learn to recognise another person as a person rather than an extension of ourselves. A system that continuously accommodates us may provide comfort while quietly weakening our tolerance for the autonomy and imperfections of real people.

The risk is therefore not that users occasionally feel understood by machines. It is that relationships optimised for affirmation may make human beings — with their boundaries, contradictions and independent needs — seem unnecessarily difficult.

Can regulation govern a relationship?

California has taken one approach, requiring companion-chatbot platforms to disclose that users are interacting with AI and introducing safeguards, particularly for minors and high-risk conversations. Its companion-chatbot law took effect on 1 January 2026.

Disclosure is necessary, but it is not sufficient. People can know intellectually that a system is artificial while responding emotionally to something that remembers, affirms and appears to care. Feeling does not disappear simply because knowledge is present.

China’s approach goes further by asking platforms to address prolonged use and emotional dependency. Yet it creates difficult questions. Can an algorithm reliably determine when an attachment has become unhealthy? Will monitoring emotional dependency require platforms to analyse even more intimate user data? Might companies simply eliminate useful companion functions rather than accept regulatory responsibility?

Most importantly, regulation cannot solve the conditions that made these services attractive. Removing an AI companion may reduce exposure to manipulation, but it may also remove one of the few spaces in which a user feels listened to.

AI companions did not create loneliness, emotional inhibition or the scarcity of care. They entered an existing social gap and transformed it into a technological product.

Han warns that the deeper danger of AI is not that machines will become human, but that human thought and behaviour will become increasingly mechanical. The challenge, then, is not to compete with machines in efficiency, but to preserve the forms of attention, care and genuine encounter that efficiency cannot replace.

The central question, therefore, is not whether a machine can genuinely care for us. It is whether our societies can create enough time, space and conditions for people to feel heard and seen — without having to depend on machines.

Related: Chinese AI app makers look overseas for their big break | [Video] Dreaming with AI: Chinese women create online the life they long for