The discussion below builds on our discussion of the draft version of these regulations, emphasizing the changes that have been made in the final version. A table showing and explaining those changes can be found below.
Major changes from the draft include:
- Coverage narrowed to address sustained emotional interactions, with specific exceptions for some applications.
- Requirement that all users provide emergency contacts.
- Allows broader use of interaction data for model training
- Shift in provisions to protect elderly users from patronizing restrictions, to empowerment
- Removes the requirement for human intervention during user emotional crises
China’s new Provisional Measures on the Administration of Human-like Interactive Artificial Intelligence Services, now jointly released by the CAC, NDRC, MIIT, and MPS, will take effect this summer, on July 15, 2026.
These Measures are an admirable attempt to get out in front of challenges that much of the world is just becoming aware of: the mental health impact of interacting regularly with non-human, virtual companions. While that sentence may read like science fiction, the issues are already pressing enough that medical professionals and policymakers are paying attention.
For better or worse, it is becoming increasingly common for people to interact with AI services that present as intelligent. That may mean asking ChatGPT for advice on navigating an awkward situation, chatting with an AI boyfriend or girlfriend, or keeping an AI service open in a browser tab as a running companion while brainstorming at work. Whatever the form, chatbots are becoming part of our social reality.
Unfortunately, these AI interactions can create risks that differ from ordinary media because AI is responsive, adaptive, and able to simulate understanding. It can build trust, shape decisions, and sustain emotionally significant interactions in ways that static content cannot, while also collecting data and personalizing influence. Those features take the risk beyond false output and into dependency, privacy intrusion, and the quiet displacement of a user’s own judgment.
Perhaps in recognition of the challenge for any legal system to keep abreast of changing technology, China has labeled these Measures as ‘provisional’, but the new rules reflect one of the most complete attempts to regulate this area.
- Scope and terminology
The draft version of these regulations defined the covered technology as:
- AI Products and services,
- That simulate human traits such as personality, thinking, and communication,
- To interact with humans emotionally,
- Through text, images, audio, video, or other means.
This broad definition left questions as to what technologies were included, and the final version has made a number of key changes to narrow the scope, although it is still far from clear:
1.“Sustained” Emotional Interaction. Although the rules still don’t define the emotional interaction requirement, it is now clearer that the Measures’ central concern is ongoing emotional engagement that can lead to emotional dependence and addiction.
The use of “sustained” to qualify the emotional interactions closely resembles California’s SB 243, which defines “companion chatbots” as systems able to “sustain a relationship across multiple interactions,” and may be intended in a similar sense. This would likely make the ability to ‘remember’ and ‘learn’ between sessions the key factor in determining coverage.
Some have suggested that the rules are meant to encompass only services specifically aimed at providing emotional services or eliciting deep emotional responses, but there isn’t much to support that. For example, as in the draft, digital relatives and companions (伴侣- a term encompassing ‘significant others’) are mentioned as a subcategory of the covered services. More importantly, the types of harms addressed aren’t limited to such services. Self-harm, extreme emotional reactions, and obsessive dependence have been associated with even mainstream LLMs outside of any role-playing function.
The better understanding is the emphasis on ‘sustained relationships’ rather than trying to quantify emotional engagement.
2, Specific Excluded Uses. Again echoing California’s SB 243, the Chinese rules narrow the category of covered AI by creating express carve-outs for certain uses. They do not apply to services such as customer service, informational Q&A, work assistance, study and education, or scientific research, which do not involve sustained emotional interaction. This loosely tracks the first set of exceptions in the California bill, although that bill clarifies that the exception applies where the service is “only” used for the listed purposes, helping resolve the issue of general-use chatbots, which may sometimes have these functions.
| Human-like AI Provisions, Article 2 | CA SB 243, b(2)(A) |
| 1. customer service 2. informational Q&A 3. work assistance 4. study and education 5. or scientific research | 1. customer service 2. a business’s operational purposes 3. Productivity and analysis related to source information 4. internal research 5. or technical assistance. |
The Chinese rules also do not incorporate SB 243’s additional exclusions for bots in video games that engage solely on game-related topics and do not involve sexually explicit conduct or mental health issues, or for standalone devices such as smart speakers that function as personal assistants. Under China’s new rules, one would have to consider whether a specific service amounted to “sustained emotional interaction”.
3. Only services and not products mentioned. This is likely less concerned with the scope of coverage than it reflects a somewhat arcane, formalist (but not trivial) distinction in tort law. In China as in many countries, harm caused by a product can invoke special non-fault liability, while recovering from services generally requires a showing of fault. Put simply, classifying AI chatbot companions as services makes private litigation less attractive and, in some cases, more difficult. That fits China’s broader regulatory approach in this area, which emphasizes ex ante compliance obligations imposed on service providers rather than ex post accountability through individual lawsuits.
II． Protections
General Requirements:
As in the draft, the Providers of human-like AI services have the primary responsibility for the safety and security of their services. (Art. 9) They are responsible for security throughout the entire lifecycle of the products and services, from design to implementation, and including all updates and revisions. Updates in security and functions are to be concurrently updated, so that there is no lag in protection for new features. (article 10)
The ‘security’ mentioned here is wide-ranging and touches on other laws to include: data security, pre-publication content reviews, and cybersecurity. Some categories listed in the draft as basic security issues, including personal information protection, and fraud prevention were removed in the final version from this, but are still addressed throughout the law (Art. 9).
Special security assessments must be conducted and reported to provincial authorities when services introduce new functions or cause major changes to existing functions, or have high user volumes (1M or more registered/ 100,000 or more monthly active users). Security assessments may also be triggered by a possible impact on national security or the public interest. (articles 22,23).
While this section was largely unchanged from the draft, there are a few notable changes. First, provincial-level cybersecurity departments are empowered to notify service providers that they should conduct an assessment. Second, potential harms to the rights and interests of individuals or organizations, rather than to broader public interests, are no longer a trigger for assessments.
1. Prohibited Service Conduct
The basic list of prohibited service conduct has been moved to Article 8, and significantly revised since the draft.
Section 8(1) now consolidates restricted content categories, largely incorporating China’s broader online content controls into this new context. This includes content endangering national security, harming the national honor and interests, conducting illegal religious activities, disrupting order, and undermining ethnic unity, as well as promoting gambling, obscenity, encouraging crimes, and conduct that impacts others’ rights, such as defamation.
One interesting note is that content going against the Core Socialist Values (a ubiquitous morality and civics campaign) is prohibited. While content supporting the Core Socialist Values is generally encouraged in China’s content control regime, this feels like an increased level of protection.
Article 8(5) and (6) more clearly address mental health impacts, AI’s tendency towards of ingratiating sycophancy is specifically identified as a potential threat to users’ social relationships, replacing the narrower and more objective standard of ‘providing false information’ in the draft. The threat of inducing ‘unreasonable’ or irrational decision making remains from the draft, but the prohibition now addresses ‘emotional’ rather than ‘algorithmic’ manipulation.
A new section at article 8(4) is added to directly address special content restrictions for minors, incorporating by reference the Measures for Categorizing Online Information that Might Impact Minors’ Physical and Psychological Health.
Prohibitions on content that encourages or glorifies self-harm, insults personal dignity, or endangers confidential information such as state secrets, work secrets, trade secrets, and personal information remain in place (Art. 8(2),(3)).
2. Mental Health Impact.
The final version of the rules remains most concerned with excessive dependency or addiction to the services and extreme emotional responses or mental health emergencies, such as risks of suicide and self-harm (but notably not harm to others). (art.9)
To prevent these risks, a number of clear preventative and response measures are required. For example, Article 18 requires conspicuous notices that the chat is with an AI and not a human being in addition to China’s normal AI labeling regime. Similarly, reminders to take a break must also be displayed after every 2 hours of continuous use.
More dynamic reminders, such as pop-up window alerts, are required when users are identified as at risk of over-reliance, although this is a vaguer standard. Curiously, the draft had required an initial registration and the beginning of new sessions, but this relatively uninvasive requirement has been removed.
Interestingly, the risks involved are not limited to threats to health, and include even risks to property. (article13). This language has been slightly qualified to include only major property risks to users that are currently occurring or have already happened.
Risk Identification and Response
The rules maintain the draft’s approach of having escalating levels of intervention.
Where a “tendency” or ‘inclination’ towards addiction or dependency is detected, for example, pop-up windows and other more aggressive means are to be used to remind users that they are chatting with a non-human, etc. (art. 18). When any user seeks to exit the service, it must be convenient and immediate, with no further interaction that might impede the users’ exit. (Art. 19).
Where users exhibit extreme emotional states, the service is to offer users consolation or comfort and encourage the user to seek help. (A requirement of having prepared templates was dropped from the draft). Where the situation escalates to clear expressions of intent to harm themselves, or threats of major damage, the necessary measures may be taken, such as providing assistance information or reaching out to the user’s emergency contact person (art. 13). (Emergency contact person information was previously only requested from the elderly, but this now applies to all users).
A proposal in the draft to require human intervention when users are in crises has been removed, but such an intervention could still be included in the mentioned ‘necessary measures.’
Serious questions still remain regarding how risks are identified. The rules still do not clarify if this process should be made through human or machine review, and either approach raises serious concerns. Human review invites intrusions on privacy, while machine review requires relying on the algorithm that is the source of the problem.
Special protection for vulnerable populations
Companionship for the elderly, childcare, and meeting the needs of special groups are among the encouraged future uses for human-like AI, and the rules also introduce special protections for such groups. (Art. 6).
Some of the most drastic changes from the draft involve elderly users. The draft adopted a paternalistic approach, prohibiting their use of services that roleplayed as their family members. These prohibitions, presumably meant to protect them from fraud or confusion, have been removed.
As mentioned above, the requirement that only elderly users provide an emergency contact person has also now been expanded to all users, acknowledging that there is no reason to portray the elderly as lacking competency.
The treatment of the elderly has been reframed as emphasizing support rather than special requirements and prohibitions. Additional services and assistance are to be made available as needed. (Art. 15).
Changes to the protections for Minors are less sweeping, but mainly because the existing law on the protection of minors is already more established. As mentioned above, rules on content harmful to minors have now been incorporated as well, in addition to special rules on the protection of minors’ personal information and general protection of minors online.
Some notable changes do address concerns raised regarding the draft, such as recognizing that minors of different ages have different needs. Article 14 now follows the general strategy for China’s requirement of mobile internet minors modes to focus protections on specific groups while empowering guardians to make informed individual decisions about usage time and spending etc.
A new ban on providing all minors with virtual family or companions (伴侣) has also been introduced, in accordance with a recent draft on regulating human-appearing characters. (Art 14). The draft had only required guardian consent for companions, which is now required for minors under 14 to access any covered services.
Identification of Minors
Outside of China, identifying minors online is a contentious issue, as it requires all users, including adults, to verify their age. Increasingly, this means uploading identification documents, creating an increased risk of harmful data breaches, and making true anonymity online impossible.
China’s internet already operates on a real-name registration system, effectively requiring all users to link their national ID to most online accounts. Even this drastic measure has not proved fully effective in blocking minors’ access to improper online content; however, as kids have learned to borrow parental accounts (with and without permission), rent other adults’ accounts, and fool ID systems. (see eg: https://finance.sina.com.cn/tech/digi/2025-12-18/doc-inhcfepi6674554.shtml)
The draft human-like AI rules seem to demand additional age verification, requiring that service providers have the capacity to identify minors. (art. 14). This could include algorithmic detection, as was hinted at previously in the Ministry of Culture and Tourism’s Opinions on Strengthening the Protection of Minors in the Online Cultural Market (at 4).
Training Data and User History Information
The general data protection regime introduced in the draft was drawn from existing law, and remains in place at Article 11(1), 16, etc.)
The final paragraph of Article 15 contains a small but meaningful change to rules on service providers’ use of interaction data. Independent consent was previously required by the draft when user interaction data OR sensitive personal information was used in model training. The final rules require independent consent only where the user information IS sensitive personal information.
Sensitive personal information is a defined category in the Personal Information Protection Law, referring to information that, if leaked, could easily cause injury. Under the updated language, most interaction data, including personal information that is merely identifying, may be used for training without independent consent. Instead, it must only comply with general notice and consent rules under the PIPL.
The table below indicates changes between the draft and finalized versions of the Provisional Measures on the Administration of Human-like Interactive Artificial Intelligence Services
Red text indicates a deletion
Green Text indicates an addition
Blue Text indicates a style or word choice shift with little change in substance.
The Final column on the right contains my notes as to the importance of the change.
| The most significant changes are noted with a gentle yellow background. |
Loading...
Reload document
Open in new tab 
Jeremy Daum is a Senior Fellow of the Yale Law School Paul Tsai China Center, based in Beijing, with over a decade of experience working in China on collaborative legal reform projects. His principal research focus is criminal procedure law, with a particular emphasis on protections of vulnerable populations such as juveniles and the mentally ill in the criminal justice system, and is also an authority on China’s ‘Social Credit System’. Jeremy has spoken about these issues at universities throughout China and in the U.S.; and has co-authored a book on U.S. Capital Punishment Jurisprudence for Chinese readers. He is also the founder and contributing editor of the collaborative translation and commentary site Chinalawtranslate.com, dedicated to improving mutual understanding between legal professionals in China and abroad.
He translates, writes, edits, does web-design, graphic design, billing, tech support, and social media outreach for China Law Translate.
- Jeremy Daum
- Jeremy Daum
- Jeremy Daum
- Jeremy Daum
English
中文(简体)
Be First to Comment