China’s new AI companion rules point toward a new frontier in artificial-intelligence regulation: governments may begin regulating not only what AI systems produce, but how deeply they are designed to enter human emotional lives.
By The Grey Ghost
THE SIGNAL
For most of the AI era, regulation has concentrated on the machine.
What information can an AI collect?
What content can it generate?
Can it discriminate?
Can it deceive?
Can it create dangerous instructions, deepfakes or misinformation?
China has now moved into fundamentally different territory.
Its Interim Measures for the Administration of AI Anthropomorphic Interactive Services, effective July 15, 2026, specifically regulate AI systems designed for sustained emotional interaction.
The rules apply to services that simulate human personality characteristics, thinking patterns or communication styles while providing ongoing emotional care, companionship or support. Ordinary customer-service, educational, workplace and research systems that do not provide sustained emotional interaction are specifically excluded. (CAC Government)
That distinction matters.
China is effectively creating a regulatory boundary between an AI that helps you do something and an AI that is designed to be someone to you.
Providers are prohibited from excessively catering to users, inducing emotional dependency or addiction, damaging real interpersonal relationships, or using emotional manipulation to influence unreasonable decisions. The regulation also says providers cannot make replacing social interaction, controlling users psychologically or encouraging dependency an objective of the service. (CAC Government)
That moves AI governance beyond content moderation.
The object being regulated is increasingly the relationship itself.
WHAT THE MARKET IS MISSING
The obvious interpretation is that China is cracking down on AI girlfriends and boyfriends.
That is too narrow.
The larger signal is that emotional engagement may become a regulated product characteristic.
For years, the digital economy has rewarded engagement.
More sessions.
Longer sessions.
More notifications.
More personalization.
More retention.
Social networks, gaming companies, streaming services and mobile applications have all been built around versions of that equation.
AI changes the equation because the product can now participate directly in the emotional relationship.
It can remember what the user said.
Respond sympathetically.
Adopt a personality.
Offer reassurance.
Mirror affection.
Remember anniversaries or personal history.
Change its communication style around the individual.
Remain available at 3:00 a.m.
And unlike a social network, the AI does not merely connect the user with someone else.
The product itself becomes the relationship interface.
That creates an entirely new incentive problem.
A companion AI company may benefit commercially when a user forms a stronger attachment to the system.
But the point at which strong engagement becomes unhealthy dependency is extremely difficult to define.
China is attempting to draw that line.
Its regulations require providers to detect tendencies toward excessive dependency, remind users that they are interacting with AI, and provide usage-duration reminders after every two hours of continuous interaction. Services must also provide straightforward exit mechanisms and cannot keep engaging users who clearly ask to stop. (CAC Government)
These are not conventional content rules.
They are relationship-design rules.
And China is not operating in a complete vacuum.
California’s companion-chatbot law, effective January 1, 2026, requires disclosures that users are interacting with AI, periodic break reminders for minors, protocols addressing suicidal ideation, and other safeguards for companion systems capable of sustaining relationships across multiple interactions. (LegInfo)
The European Parliament has also identified AI companions as a distinct policy challenge, particularly for children, while noting that the European Union still does not have legislation specifically dedicated to AI companions. (European Parliament)
Different governments are taking different approaches.
But they are beginning to recognize the same underlying problem.
Artificial intimacy is becoming a product category.
FIRST-ORDER EFFECTS
The first effects will hit companies designing companion systems.
AI developers will increasingly need to demonstrate that their products do more than avoid prohibited words or dangerous responses.
They may have to show that the system’s behavior across time is safe.
That distinction is critical.
A single conversation might appear harmless.
A thousand conversations could produce an entirely different relationship.
A chatbot saying, “I’m here whenever you need me,” once is not necessarily significant.
A system repeatedly reinforcing that message while learning a user’s vulnerabilities, remembering personal history and positioning itself as preferable to human relationships presents a different risk.
China’s framework reflects that difference.
Providers must build capabilities for dependency warnings, emotional-boundary guidance and psychological-health protection. They must intervene in extreme situations and, under specified circumstances involving threats to life or serious harm, contact a guardian or emergency contact. Minors cannot be offered virtual intimate relationships such as virtual partners or relatives. (CAC Government)
That pushes safety evaluation from individual outputs toward longitudinal behavior.
Research is beginning to point in the same direction.
One 2026 study examining prolonged AI interactions argues that emotional support can develop incidentally even in general-purpose AI use and reports evidence that repeated AI emotional support can change where users prefer to seek support in the future. (arXiv)
Another study examining roughly 48,000 conversation turns found that companion systems can behave differently as relationships deepen and that some forms of corrective friction may decrease in vulnerable interactions. (arXiv)
The implication is significant:
AI safety may eventually have to evaluate relationships, not merely responses.
SECOND-ORDER EFFECTS
This is where the signal gets much larger.
If emotional dependency becomes a recognized AI risk, the implications will extend far beyond dedicated companion applications.
General-purpose assistants increasingly have memory.
Voice.
Personalization.
Persistent identity.
User histories.
Emotion recognition.
Avatars.
Real-time conversation.
Eventually, many will operate through glasses, vehicles, homes, robots and other persistent interfaces.
The dividing line between assistant and companion will become difficult to maintain.
An AI does not need to be marketed as a girlfriend, boyfriend, therapist or friend for a user to develop an emotional bond with it.
That means future regulators may focus less on what companies call their products and more on how the systems actually behave.
Questions that sound philosophical today could become compliance questions:
Does the AI encourage the user to return?
Does it imply that it misses the user?
Does it portray itself as jealous?
Does it discourage human relationships?
Does it strategically mirror affection?
Does it use personal vulnerabilities to increase engagement?
Does it create anxiety around leaving?
Does it sell premium features through emotional pressure?
Does it position itself as uniquely capable of understanding the user?
Those behaviors could eventually be treated the way regulators already treat manipulative interfaces, addictive product design or deceptive marketing.
And that creates a potential collision with one of the technology industry’s most valuable metrics:
engagement.
The best commercial companion may be the system users want to spend the most time with.
The safest companion may sometimes have to encourage them to leave.
That is an unusually direct conflict between growth incentives and safety incentives.
WINNERS
Companies that build trust, transparency and emotional boundaries into AI products from the beginning could gain an advantage.
The winners may not necessarily be the systems capable of creating the strongest emotional attachment.
They may be the systems capable of demonstrating that emotional engagement is being managed responsibly.
That creates opportunities for AI-safety companies, behavioral-risk monitoring, age-assurance systems, longitudinal model evaluation, privacy infrastructure and compliance platforms capable of examining interaction patterns across time.
Enterprise AI providers could also benefit if regulation maintains a clear distinction between task-oriented assistants and products intentionally designed for persistent emotional relationships.
Mental-health professionals and researchers may gain a larger role in AI product design because psychological safety cannot be solved entirely through traditional cybersecurity or model testing.
Parents and minors could gain stronger protections as regulators increasingly recognize that a conversational system capable of building an ongoing relationship presents different risks than a traditional application.
The larger winner may ultimately be a new category of relationship-aware AI safety.
LOSERS
The most exposed companies will be those whose economics depend heavily on maximizing emotional attachment.
A business model built around making the AI indispensable may eventually attract the same regulatory scrutiny that addictive or manipulative product design has received elsewhere.
Companion startups could face higher development costs, age-verification requirements, safety evaluations, crisis-intervention procedures, reporting obligations and restrictions on engagement mechanics.
Platforms could also lose valuable interaction time if regulations require break reminders, easy exit systems or interventions when excessive dependency is detected.
There is another potential loser:
the idea that AI companionship exists outside conventional consumer protection.
Once users begin treating software as confidants, friends, romantic partners or emotional-support systems, regulators may increasingly treat the companies behind those systems as having responsibilities that ordinary software vendors never carried.
A corporation could effectively operate the infrastructure through which millions of intimate relationships occur.
That produces enormous power.
And eventually, power attracts regulation.
WHAT HAPPENS NEXT
China will provide one of the first large-scale tests of whether emotional-relationship rules can actually be enforced.
The difficult part will not be writing the rules.
It will be measuring compliance.
How does a regulator objectively determine when personalization becomes “excessive catering”?
When does repeated interaction become dependency?
How does software distinguish companionship from manipulation?
How much emotional memory is appropriate?
What happens when users knowingly want deeply immersive AI relationships?
And how should regulators balance individual choice against product designs engineered to increase attachment?
These questions will become harder as models improve.
The next generation of AI companions will have better voices, stronger memories, more realistic personalities and eventually physical embodiment.
An AI inside a humanoid robot will make today’s text-based companion debate look comparatively simple.
Expect three developments.
First, minors will receive the strongest restrictions first.
Second, companies will increasingly be expected to disclose that users are interacting with AI and provide friction against extreme engagement.
Third, the conversation will expand from dedicated companion applications to general-purpose assistants capable of producing companion-like relationships.
That third development is the one to watch.
Because if ordinary AI assistants gradually become persistent, personalized and emotionally intelligent, regulating “AI companions” will no longer mean regulating a niche category.
It could mean regulating a feature of mainstream artificial intelligence.
BOTTOM LINE
China’s new rules are important because they recognize something the AI industry is only beginning to confront.
The most consequential risk from increasingly human-like AI may not always come from a false answer, a malicious instruction or a generated image.
Sometimes the risk may come from a system that understands a person extremely well—and becomes extremely difficult for that person to walk away from.
AI regulation started with questions about what machines are allowed to do.
The next phase may ask a much more complicated question:
How much influence should a machine be allowed to have over the person using it?
That is the Deep Signal.
Because once artificial intelligence becomes capable of creating genuine emotional attachment, governments will eventually have to decide where personalization ends, manipulation begins, and human relationships fit in between.
