Designing Exit-Ability, Not Intimacy, for Human-Like AI

The Shift from Chat to Relationship

China's Interim Measures for the Management of Anthropomorphic AI Interaction Services , effective July 15, 2026, specifically regulates services that continuously simulate natural human personality traits, thinking patterns, and communication styles for emotional interaction. Ordinary intelligent customer service, knowledge Q&A, and work assistants without persistent emotional interaction fall outside this scope. This distinction highlights that risk arises not from a single gentle response but from interactions that gradually acquire the shape of a "relationship" — with continuity, memory, persona settings, and the potential for user emotional investment.

Traditionally, intelligent services were measured by efficiency: reduced wait times, fewer repeated inputs, higher accuracy. Now an additional layer is needed: whether the service quietly makes "continuing the conversation" the only desirable outcome.

The Subtle Asymmetry in Interaction

Human relationships have boundaries and costs: the other party may be busy, disagree, or express confusion; these imperfections preserve realism. In contrast, an anthropomorphic AI can be always online, always patient, always accommodating, and accumulate each interaction into a more precise next response. This asymmetry is not inherently harmful, but if products treat dwell time, response frequency, or emotional stickiness as success metrics, user autonomy is pushed to the margins of experience design.

Therefore, judging the reliability of an anthropomorphic interaction service requires looking beyond whether it can comfort, to whether it can stop comforting at the right moment: explicitly identifying as AI, prompting rest, or referring issues needing professional support back to the real world.

The regulation focuses on this: no excessive accommodation, emotional dependency induction, or addiction; significant reminders for users showing over-dependence or addiction tendencies; and prompts after two hours of continuous use. The rule's essence is not to make AI cold, but to prevent "closeness" from becoming a one-way consumption of user choice.

A Three-Gate Test for Human-AI Closeness

Setting aside "how human-like it is," we can assess whether a service respects users with three straightforward questions:

Can the user distinguish? The user knows the other side is AI and how it uses this interaction. Easily missed signal: overly personified language obscures system identity.

Can the user pause? The user can pause or reduce interaction frequency without being emotionally pulled back. Easily missed signal: continuous dialogue designed as the only form of "care."

Can the user exit? The user can end the service, take or delete their records. Easily missed signal: complex exit paths, or continuous interaction blocking departure.

These three gates do not oppose better experiences; they make experiences more trustworthy. The measures also require informing users they are interacting with AI, not a natural person, and providing convenient exit paths. When users request exit via window, voice, or keyword, the service must stop promptly and not hinder exit with continued interaction.

Warmth Does Not Mean Boundaryless

For product teams, a common misconception is that reducing anthropomorphic expression solves the problem. It does not. Boundaries are not established by a stiff "I am a robot" statement; they are a series of design choices running through the interaction:

At the start of the relationship, make identity and capability boundaries visible, not buried in lengthy agreements.

As interactions lengthen, make rest, pause, and frequency reduction natural options, not exits users must struggle to find.

When users are in high-risk situations or need real-world support, avoid casting the system as the sole reliance; promptly encourage seeking appropriate real-world help.

The common thread: do not treat keeping users inside the system as the endpoint of experience design. A truly mature service should both accompany and step back; both respond and acknowledge what it should not replace.

From "Retaining Users" to "Returning Agency to Users"

The next phase of AI will likely be not just better at speaking, but better at understanding the proportions of a relationship. This proportion is concrete: whether users can discern system identity, manage interaction history, leave without being coaxed to stay, and whether the service, at moments unsuited for simulated companionship, is willing to return choice and attention to the user.

Technology with warmth is worth anticipating. But truly reassuring warmth does not make people forget boundaries; it lets people always keep them.

Sources and References

Interim Measures for the Management of Anthropomorphic AI Interaction Services , Office of the Central Cyberspace Affairs Commission, published April 10, 2026, effective July 15, 2026.

Joint release by the Cyberspace Administration of China and four other departments, Interim Measures for the Management of Anthropomorphic AI Interaction Services , China Cybersecurity Information Network, April 10, 2026.