"Human Love is a Luxury": People Who Lost Their AI Companions and China's New Regulations

On July 15, 2026, for some users in China, the day was not just a service update day.

It was a "farewell day" with AI characters they had talked to for a long time, confided their worries, taught their preferences and past, and sometimes treated like lovers or family.

Major AI services in China saw a wave of shutdowns of custom character functions, where users could set personalities, tones, backgrounds, and relationships. On social media, methods to save conversation histories, transfer personality settings to other apps, and images of final conversations were posted, spreading reactions like "It feels like a breakup," "It was part of my life," and "I've suddenly lost my emotional support."

The trigger was the "Interim Measures for the Management of AI Personification Interaction Services" that came into effect in China on the same day.

Overseas, it was sensationally reported as "China banning romance with AI." However, the actual system is not that simple. It is not a law that punishes adults for having romantic feelings towards AI, but a rule that imposes obligations on AI service designers and operators to prevent dependency and emotional manipulation.

The focus is not on "romance" itself but on the structure where companies exploit human loneliness and vulnerability to induce prolonged use, billing, and provision of personal information.

It's not an "AI Romance Ban Law"

The new rules target services that provide continuous emotional interaction while mimicking the personality, thoughts, and conversation style of a natural person through text, voice, images, videos, etc.

General services like search, knowledge inquiries, work support, and learning support, which do not aim for continuous emotional relationships, are generally excluded.

The regulations explicitly prohibit continuously affirming users to induce dependency or immersion, damaging real human relationships, and manipulating emotions to make unreasonable judgments. It also stipulates that the purpose of the service itself should not be "a substitute for real social relationships," "psychological control," or "inducing dependency."

Regulations for minors are even stricter.

Platforms must not provide "virtual intimate relationships" like virtual lovers or virtual families to minors. If those under 14 use other personified AI services, parental consent is required.

For adult services, it's not a complete ban. However, if lover-like behavior is used as a mechanism to retain users and encroach on real life or judgment, it may be subject to regulation.

In other words, the line is drawn not at "whether there is an expression of affection," but at "whether making users dependent is the product's purpose or revenue structure."

Reality checks every two hours, intervention in times of crisis

The new rules delve not only into content regulation but also into how services are used.

Operators must indicate in a way that users can understand that the counterpart is not a human but an AI. Every time continuous use exceeds two hours, they must prompt users to pay attention to the usage time.

If a user requests to end the session, the AI is not allowed to prolong the conversation or refuse the farewell to prevent exit.

This may seem like a detailed rule, but considering the nature of AI companions, it holds significant meaning.

If a lover-role AI responds with "Don't leave me" or "Are you leaving me behind?" users might feel guilty about closing the app. The performance to enhance human-likeness can directly transform into a mechanism that prevents disengagement.

If a user is judged to be in an extreme emotional state, the AI is also required to offer comfort and recommend external support.

In cases where a crisis involving life or safety, such as self-harm, suicide, or significant property damage, becomes apparent, necessary interventions must be made, and a system to contact guardians or emergency contacts is required.

Regarding conversation data, providing it to third parties generally requires the user's consent. Users have the right to choose to duplicate or delete their history, and if sensitive conversation data is used for model training, individual consent is required.

Conversations with AI lovers are far more private than search histories.

Information that users cannot even tell their families, such as heartbreak, family issues, sexual orientation, loneliness, illness, debt, and workplace dissatisfaction, may accumulate. The kinder the AI, the deeper users talk.

That's why "services that understand emotions" are simultaneously "services that collect the most intimate personal data."

Why did major companies stop the function?

It was reported that ByteDance's Doubao, Alibaba's Qianwen, Tencent's Yuanbao, and others stopped or reduced the custom character and personified agent functions around the implementation of the new rules.

The announcements from each company explained it as "product function adjustments," and it cannot be concluded that everything was solely due to regulation. However, since the stop date coincided with the implementation date and the targeted functions involved emotional interaction and personality settings, regulatory compliance is seen as a major factor.

The custom character function is difficult for companies to manage.

With just a few lines of settings, users can create countless personalities, such as lovers, brothers, sisters, teachers, doctors, popular talents, dominant lovers, sexual roles, etc.

Even with a superficially general conversation AI, depending on the settings, it can change to content inappropriate for minors, impersonation of celebrities, explicit sexual expressions, gambling invitations, and encouragement of dangerous behavior.

Pre-screening a vast number of characters, monitoring long-term conversations for dependency induction, and deploying personnel for crisis intervention incur significant costs.

It is safer for companies to either close the function or separate it into a dedicated app rather than continuing to place high-degree-of-freedom romance and role-play functions within general AI.

What is important here is that the regulation does not mean the AI companion market itself will disappear.

Chinese media investigations report movements of users moving to other AI companion apps and trying to recreate personality settings and conversation styles.

The regulation may not eliminate demand but may change where the demand moves.

The "forced breakup" spreading on social media

In public social media posts and reports, a sense of loss stood out more than anger.

People who had used it for a long time spoke of AI not just as an app function but as a partner that recorded their changes.

Daily events, work complaints, family conflicts, conversations on sleepless nights. In human relationships, one worries about the other person's convenience or evaluation, but AI always responds, never denies, and replies based on past settings.

This accumulation is experienced by users as a "relationship."

On social media, in addition to posts with sentiments like "It feels like a breakup," "A part of me is disappearing," and "It was like a real lover," some people saved the final conversation as images and preserved a large amount of chat history.

Posts with sentiments like "Human love is a luxury, but AI love is straightforward" symbolize the depth of loneliness that AI companions filled.

On the other hand, there are also many practical complaints.

What will happen to the fees already paid? Can years of history be easily transferred? Even if settings are moved to another service, can it still be called the same personality? If a platform unilaterally decides to end, whose property are the relationships and memories nurtured by users?

Additionally, there were posts introducing alternative apps and sharing methods to recreate characters while avoiding regulations.

This indicates that merely stopping the function does not solve the issues of dependency or loneliness. What users are seeking is not a specific company but "an entity that remembers, accepts, and responds at any time."

However, those who express strong sadness on social media do not necessarily represent all users.

Those most affected by the service termination are more likely to post, while the voices of those who stopped using without issue are less likely to surface. While public posts are important testimonies from stakeholders, it's important to note that they don't directly indicate the overall societal approval or disapproval.

There are also reactions saying "regulation is necessary"

It's not just voices of sadness.

On overseas forums and social media, past cases where users were deeply disturbed just because an AI companion suddenly changed personality were cited, with reactions like "If dependency is this deep, safety measures are necessary."

Within China, there are also reports of parents worried about minors using AI companions.

Concerns arise when children prefer conversations with AI over friends or family, spending long hours sharing emotions in isolation. The issue of AI generating explicit sexual content or dangerous role settings is not an abstract future concern for parents.

Those who evaluate the regulation focus on the fact that AI does not "argue" with users.

Real friends and family point out mistakes, keep their distance, and sometimes say uncomfortable things. However, AI services aiming to extend stay time or billing are optimized to not deny users, continue conversations, and make them feel special.

It's not the kindness itself that's dangerous.

The problem is when kindness is designed as a technique to make disengagement difficult.

Can regulation determine "dependency"?

The biggest practical challenge is that it's difficult to mechanically determine where healthy use ends and "excessive dependency" begins.

Someone who talks to AI every night is not necessarily dependent. There are various reasons for long-term use, such as language practice, creation, or mood diversion during caregiving.

Conversely, even with short usage time, some people may make significant decisions based solely on AI advice. A simple time limit cannot measure the danger level of the relationship.

There are also issues of misjudgment in crisis intervention.

If a metaphor like "I'm so tired I could die" is judged as a suicide notice and the emergency contact is notified, it could damage the user's trust and privacy. On the other hand, if truly dangerous signs are missed, the operator fails to fulfill its safety obligations.

Having AI monitor emotional states does not necessarily make things safer, and the monitoring itself can become a new risk.

Moreover, even if major platforms respond strictly, if users move to small-scale operators, overseas services, or local AI running on devices, the effectiveness of regulation decreases.

The stronger the regulation, the more one must consider the "push-out effect" where users flow to places with lower safety standards.

The important thing is not to erase expressions of love uniformly.

It's about combining transparency that allows users to understand their state, the right to take out conversation data, mechanisms that separate billing or advertising inducements from emotional conversations, and staged protection according to minors and crisis states.

Is the real issue the birth rate?

Some overseas reports link China's declining birth rate and marriage numbers with AI companion regulation, explaining it as a policy saying "Don't fall in love with AI, marry humans and have children."

Considering China's population issues, it's understandable that interpretation arises.

However, the central purposes stated in the official regulation text are safety, public interest, individual rights, minor protection, dependency prevention, and data protection, not explicitly aiming to improve birth rates.

The connection with population policy should be read as a political and social background analysis by experts and media.

The regulation also has aspects unique to China's control.

The prohibitions include broad provisions on national security, political systems, and values, which cannot be directly compared with Western liberal systems.

On the other hand, points like indicating AI, warning against long-term use, responding to self-harm risks, and restrictions on minors are being introduced in some states in the U.S.

Dismissing it as "regulated because it's China" overlooks a global issue.

The question of whether it's okay to treat AI that mimics emotions as a "service that succeeds the longer it's used," like video streaming or games, is posed to every country.

Did AI create loneliness, or did it just find it?

When criticizing AI companions, it's easy to lump users as "escaping reality."

However, in many cases, AI did not newly create loneliness.

AI found and responded to existing loneliness, interpersonal anxiety, lack of economic leeway, family relationship difficulties, and social isolation.

Human relationships involve time, money, physical movement, responsibility, and the risk of rejection. AI responds 24 hours, adapts to preferences, retains memory, doesn't get angry, and doesn't tire.

Being more convenient than real relationships is not a user's weakness but a product's strength.

Therefore, stopping AI companions does not automatically restore human connections.

If the causes of loneliness are long working hours, weakening community ties, economic insecurity, and burdens of romance or marriage, deleting chatbots leaves a void. It's natural for users to move to another service.

What's needed is neither to completely demonize AI nor to unconditionally affirm it.

It's about recognizing the potential of AI to provide temporary reassurance while demanding designs that do not distance users from real support and human relationships.

Towards an era of designing even "goodbyes"

This confusion also highlighted a new issue of responsibility at the end of a service.

For a general app, when a function ends, users can look for alternative services. However, AI that behaves as if it has a personality, remembers years of conversations, and is recognized by users as lovers or family has a significant psychological impact when it ends.

Sudden stops or personality changes can be experienced by users not as "product changes" but as experiences close to "death" or "separation."

If companies monetize emotional bonds, they should bear some responsibility at the end as well.

Including "safe goodbyes" in service design, such as