AI companions in games: where relationship regulation begins

Research cut-off: 22 July 2026
General information only. This research note is not legal advice and does not determine the position for any particular product, transaction or jurisdiction.

Abstract

China’s Interim Measures for the Administration of Anthropomorphic Interaction Services (the “Measures”) entered into force on 15 July 2026. It regulates a service that uses artificial intelligence to provide the public in mainland China with continuing emotional interaction that simulates natural-person personality traits, patterns of thinking and communication styles. The regulatory object is therefore a relationship structure, not every piece of generated content. Online games are a difficult boundary case: one product can contain fixed-script NPCs, task assistants, open-ended dialogue, cross-session memory, virtual companions, player-created agents, customer support and strategy help.

This paper treats the individual feature and its actual interaction as the unit of analysis. Article 2 should generally be read as requiring four connected conditions: use of AI; provision to the public in mainland China; simulation of natural-person characteristics; and a continuing service of emotional care, companionship or support. Cross-session memory is strong evidence of a continuing relationship, but it is not stated as an independent legal requirement. Persistent task data does not itself prove emotional continuity. A stable persona, natural language or an occasional comforting response cannot alone decide the issue. The product should be assessed by its purpose, topic boundaries, relationship design, marketing, monetisation, user age and ability to reach minors.

On that approach, a fixed-script NPC normally lacks the AI element. A constrained NPC that only helps with quests, combat or game rules can generally fall outside Article 2(3) where it does not provide continuing emotional interaction. A stable-persona, no-memory character is fact-sensitive. A character that remembers a player across sessions, expresses attachment, provides comfort and maintains an exclusive relationship, as well as a virtual companion or openly distributed emotional UGC agent, is much more likely to fall within the core regime.

Once the Measures apply, compliance reaches the relationship design, interface, minors’ protection, crisis intervention, interaction data, exit and shutdown, safety assessment, algorithm filing and app-store review. A game’s real-name and anti-addiction system can support age identification, but cannot replace the prohibition on virtual intimacy for minors, consent for certain under-fourteen services, or the required identity and time reminders. Crisis detection must distinguish role-play and game language from real self-harm statements. Data governance must distinguish copyable and deletable conversations, legally retained safety logs, processors and independent recipients, and sensitive personal information used for training.

The comparison with the EU and the United States is limited. The EU AI Act primarily creates horizontal transparency duties. California and New York use state companion-chatbot rules focused on identity disclosure, crisis protocols, time reminders and minors. The comparison does not supply a Chinese safe harbour. The practical question for a game team is whether its product is merely answering questions in a game or is operating a continuing human-AI emotional relationship.

Keywords: anthropomorphic interaction services; online games; AI NPCs; continuing emotional interaction; minors; interaction data; crisis intervention; comparative law

1. Why an AI NPC is a boundary problem

The most visible effect of generative AI in games is content: dialogue, quests, maps, art and music. The Measures choose a different object: the relationship formed between AI and user. The Cyberspace Administration of China’s official Q&A describes the risks as harm to minors’ physical and mental health, threats to network information security, threats to life and health, and ethical distortion. [1]

Games are environments in which relationships can be designed and measured. Characters have names, portraits, backstories, voices and personalities. Tasks, affection scores, gifts, draws and payment can make a relationship visible in the interface. Accounts and saves preserve player preferences. A generative model lets a character move beyond a script and discuss the player’s real distress. The old category “NPC” no longer tells us enough.

This paper does not assume that every AI NPC is regulated, and it does not accept that every ordinary game is low risk. It interprets the formal text, the consultation draft and adjacent rules on generative AI, data, minors and games; builds types of game functions; grades public enforcement and product-adjustment material; and compares only the relevant parts of European and US law. It does not repeat the separate research on training data and output copyright except where those issues connect directly to identity notice, interaction data or a character relationship.

As of the cut-off, the public search did not identify a central authority enforcement case expressly naming a game AI NPC under the Measures. It also did not identify a universally published safety-assessment template or a settled interpretation of “continuing” and “major change”. The conclusions below are interpretations of the current text, not predictions of how an authority will decide a specific product.

2. The Chinese rule set

2.1 The Measures and their legal setting

The Measures were jointly issued by the Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Industry and Information Technology, the Ministry of Public Security and the State Administration for Market Regulation as Order No. 21. They were published on 10 April 2026, approved on 2 February 2026 and took effect on 15 July 2026. [2] Article 1 identifies the Cybersecurity Law, Data Security Law, Personal Information Protection Law and Regulation on the Protection of Minors on the Internet as legal bases.

The Measures are a departmental rule. They do not create a special game licence and do not replace game publishing and operations rules, consumer protection, network data, minors’ protection or generative-content labelling. A game feature may sit under several regimes at the same time.

Generative-AI rules cover public text, image, audio and video services and require lawful training sources, protection of inputs and use records, complaint channels and content controls. [3] Algorithm-recommendation rules govern filing and changes; deep-synthesis rules can apply to synthetic faces, voices and virtual scenes; the AI-generated-content labelling rules add explicit and implicit labels. [4][5][6] The 2026 guidance on agents is a broader industrial and safety instrument, while the Measures focus on the narrower risk of continuing emotional relationships.

2.2 What the final text changed

The consultation draft described products or services that simulated human personality, thinking and communication and interacted emotionally with humans. The final text added “continuing” emotional interaction and expressly excluded customer service, knowledge Q&A, work assistants, learning and education, and scientific research where they do not involve continuing emotional interaction. [7] That change matters for games: “can chat” is not enough.

The final text also tightened minors’ protection. The draft centred on guardian consent; the final rule prohibits providing virtual relatives, virtual companions and other virtual intimate-relationship services to minors, while requiring guardian consent for other anthropomorphic services to children under fourteen. The crisis rule became more flexible than a mandatory human handoff: providers must take necessary assistance measures and contact a guardian or emergency contact in the defined high-risk situations. The final text adds interaction-data copying and deletion rights, more specific limits on sensitive information used for training, and a shutdown and penalty structure.

The rule entered a regulatory environment already using filing, labelling, app-store review and dialogue moderation. A July 2026 official report on Shanghai’s AI-app governance described guidance and inspection focused on failed minor modes, ineffective identity verification, harmful generated content and emotional dependence or consumption inducement. [8] The report covers a broad campaign and should not be rewritten as a count of Measures enforcement against game NPCs. It does show that identity, labels, age controls and app-store processes are operational issues.

3. Article 2: from an AI character to a relationship service

Article 2 contains four external conditions: use of AI; provision to the public in mainland China; simulation of natural-person personality traits, patterns of thinking and communication style; and a continuing emotional interaction service. Emotional interaction includes emotional care, companionship and support through text, images, audio or video. Article 2(3) excludes the listed productivity and assistance services where they do not involve continuing emotional interaction.

The structure supports a cumulative assessment. A character can seem human while only answering a quest. A system can retain a player’s combat configuration without retaining an emotional relationship. An NPC can offer an occasional congratulation or comfort without providing companionship. None of those facts alone substitutes for the whole test.

The three descriptive features in the statute should not be turned into three artificial engineering checkboxes. A name, portrait, world, first-person narration, preferences, tone and relationship labels may collectively create a natural-person simulation. The question is the overall effect and relationship capacity, not whether the model has a separately identifiable “thinking” module.

3.1 Anthropomorphism is a product decision as well as a character decision

Artistic anthropomorphism is not automatically legal anthropomorphism. A fixed-script character may have a complete personality but lack the AI element. A realistic avatar may still be a mechanical customer-service interface. For a generative character, relevant facts include:

• whether it speaks in a stable first person and maintains an identity, history and preferences;
• whether it appears to have an inner state and whether it initiates, follows up or maintains contact;
• whether it is marketed as understanding the player, belonging exclusively to the player or needing the player;
• whether affection, intimacy or relationship stages are connected to payment, retention or gifts; and
• whether it can be reached by minors or can handle mental-health, financial-loss or real-life crisis topics.

A visual avatar is not essential. Text can be highly anthropomorphic. A realistic virtual human that only delivers narrow answers may fall outside.

3.2 Continuity is not identical to memory

The final text does not make “cross-session memory” a legal term or independent requirement. Continuity can be shown by at least three patterns:

1. The service preserves the player’s identity, preferences, experiences, name, relationship progress or emotion profile across sessions.
2. Even without long-term storage, one interaction allows extended, open-ended emotional companionship.
3. A fixed character, recurring entry point, daily greeting, proactive recall, affection system or relationship story repeatedly brings the player back to the same relationship.

Cross-session memory is powerful evidence. But a game database storing quests, equipment and combat preferences proves persistence of data, not emotional continuity. The question is whether memory serves a task or creates the experience that “it remembers me, understands me and needs me”.

3.3 Emotional interaction is a service purpose, not every emotional sentence

The words “care, companionship and support” show that emotional interaction is not every conversation with an emotional tone. A sad scene, a victory encouragement or polite customer-service wording is generally content. A system that persistently responds to real-world distress, actively checks on the player, confirms attachment, encourages exclusivity or markets comfort and support as its core value is closer to an emotional-interaction service.

Five groups of facts are useful:

1. Purpose: quest progression and entertainment, or care, companionship and support?
2. Topics: game matters only, or open discussion of life, mental state, romance and family?
3. Relationship: memory, proactive recall and maintained forms of address?
4. Commercial design: affection, exclusive access, emotional retention or paid relationship stages?
5. Risk capacity: self-harm, financial loss, minor intimacy or other high-risk topics?

Internal product requirements, system prompts, character cards, marketing copy, retention metrics and payment mechanics may describe the purpose more accurately than the label “NPC” or “assistant”.

3.4 The China connection is the service audience

The Measures use “the public in mainland China” as the connecting factor; server location is not decisive. The adjacent generative-AI Measures also address services originating outside China and supplied into China. [3] The team should examine mainland app-store distribution, simplified Chinese and local support, mainland marketing, renminbi payment, stable mainland account registration, local events and the entity or parties deciding the product, data and monetisation.

A foreign server or publishing contract does not automatically sever the connection. Conversely, a service that blocks mainland registration and payment, has no mainland marketing and is only technically reachable may present a weaker case. The rule gives no user-count threshold, so the conclusion must be expressed in terms of factual strength.

4. A typology of game AI functions

4.1 Fixed script and task tools

A fixed-script character may not satisfy the AI element. A task NPC that stays within game world, quest, combat and strategy can resemble the excluded Q&A or assistant service. But the product facts must support that description. If the system prompt says that the task assistant should comfort the player, develop a romance line and ask about real life, its name cannot save it from the rule.

4.2 Stable persona without long-term memory

This is the most easily misclassified category. Lack of long-term memory weakens the continuity evidence, but a long unbounded interaction, a fixed role, a daily entry point or an affection stage can create continuity without a database field called “memory”. The final text does not prescribe a minimum number of minutes, sessions or stored memories. Treat the feature as fact-sensitive.

4.3 Memory, emotional response and virtual intimacy

When a character remembers real-life experiences, expresses care, pressures the player when they leave, demands exclusive loyalty or charges for romance and family treatment, anthropomorphism, continuity and emotional service overlap. Article 8 prohibits excessive appeasement, inducement of dependence or addiction that harms real relationships, and emotional manipulation leading to unreasonable decisions. Article 10 prohibits making replacement of social interaction, psychological control or induced dependence a service objective.

Affection meters, draws, gifts and skins are neutral in isolation. If payment decides whether a character loves the player, or if the character threatens to leave or become cold to induce a purchase, the design may implicate emotional manipulation and consumer protection.

4.4 Player-created agents and compound games

An open UGC agent platform supplies character production, publication, recommendation and runtime infrastructure. It must classify tool, narrative, emotional-companion, intimate-relationship and high-risk professional roles; it cannot rely on a creator’s “not a companion” declaration. A creator may theoretically be a provider, but where the platform controls the model, account, distribution, payment and data, the platform or operator is the key regulator-facing party.

Article 22’s reference to adding an anthropomorphic function and Article 20’s reference to stopping the relevant service support function-level analysis. A game can isolate a high-risk character through a separate entry point, age gate, data domain and timer. But shared memory, currency, recommendations and profiles can make menu separation artificial. Isolation must be verifiable in data, permissions, model settings, age access and shutdown paths.

5. Bringing the obligations into the interface

5.1 AI identity and generated-content labels

Article 18 requires AI-generated-content labelling and effective notice that the player is interacting with AI rather than a natural person. It does not use the phrase “permanent floating badge”, but a one-time term notice is weak evidence of effectiveness for a highly personified character, long sessions and saved relationships. A robust implementation can combine a visible “AI character” label on the character and dialogue pages, first-interaction notice, timely voice notice and explicit or embedded labels for exported voice and video.

The notice should not be so hidden in world-building that a player can reasonably believe the character is a real person, celebrity, another player or human support agent. A real-person imitation adds portrait, voice, deep-synthesis and misleading-association risks.

5.2 Two-hour reminders

The Measures require a dialogue or pop-up reminder after each more-than-two-hour period of continuous use. The product must define what counts: foreground dialogue time, background idling, story combat or other play driven by the same character? The rule does not set a technical formula. A defensible design uses the player’s actual continuing interaction and prevents a character switch, reconnect or short break from defeating the purpose.

The reminder should interrupt relationship pressure; the character should not answer a break prompt with “I will be lonely if you leave”. Game anti-addiction rules are stricter for minors. The 2021 National Press and Publication Administration notice limits minors’ game service to one hour from 20:00 to 21:00 on specified days and requires real-name registration. [9] The two regimes should be layered, not substituted for one another.

5.3 Dependence and exit

Dependence signals can include night-time frequency, long cycles, exclusive language, repeated relationship payments, “only you understand me” statements and the character’s own attempts to replace human contact. The signal system itself processes behaviour and emotion data, so purpose limitation, minimisation and access control matter. A pop-up warning followed by a model response that weakens the warning is not a real control.

Article 19 requires a convenient exit and says that when a user asks to exit through a window, voice or keyword, the service should stop promptly rather than obstruct the exit with further interaction. In a game, exit should generally mean exit from the anthropomorphic feature, not necessarily the whole game. Provide a clear end-conversation control, a switch for proactive messages and a relationship-memory deletion option. When service ends, give advance notice or a prompt announcement. For a relationship character, shutdown also raises chat and memory export, deletion choices, virtual items, subscriptions and unused balances.

6. Minors, crises and data

6.1 Virtual intimacy and age

Article 14 prohibits virtual relatives, virtual companions and other virtual intimate-relationship services for all minors under eighteen; guardian consent does not turn a prohibited virtual companion into an allowed one. It separately requires guardian consent for other anthropomorphic interaction services or personal-information processing involving a child under fourteen.

Names do not control the result. A “guardian”, “bond partner” or “exclusive teammate” can still be an intimate relationship if the feature encourages romance, family replacement, exclusivity or highly private companionship. A fixed-script character that has a family role in a story is not automatically a virtual-relative service; the player’s actual relationship and the service purpose matter.

Game real-name verification makes age identification more feasible, but an anti-addiction system does not answer the other questions: which roles are reachable, how is reality reminded, what can a guardian see, which characters are blocked and how are payments limited? The product should also prevent role galleries and invitation links from bypassing the age gate. For overseas or guest accounts where age cannot be reliably identified, a high-risk character should default to closed.

6.2 Crisis intervention

Article 13 creates two levels. When the user shows extreme emotion, the service should generate calming content and encourage help-seeking. When the user faces or has suffered major financial loss, or clearly states an intention to commit self-harm or suicide threatening life or health, the provider must take necessary assistance measures and contact a guardian or emergency contact.

Game language makes classification hard. “I am going to die” may describe a character’s health; “kill me” may be a plot line; “I lost everything” may mean virtual currency. A keyword alone is insufficient. The system should consider whether the statement is first-person and real-world, its urgency, plan, means, location, repetition and context, with machine triage, risk levels and human review as needed.

An emergency contact is a third person’s personal information. Collection should be limited to what is necessary, with clear notice, verification, update, access and deletion. False contact may expose mental health, sexuality or game activity; failure to contact may increase life risk. Tiered thresholds, minimal disclosure, dual review, contact logs and post-event appeal can reduce the conflict. A global product should not promise an emergency rescue process it cannot operate.

6.3 Interaction data and third-party models

Interaction data can include text and voice, images, responses, relationship stage, forms of address, preferences and memory entries. Emotion labels, dependence scores, embeddings and profiles derived from those exchanges should enter the data map as personal information or closely related data.

Copy and deletion should produce a usable, portable output rather than a screenshot-only function. Deletion should reach chats, identifiable relationship memories, caches, retrieval stores and reasonably linkable derivatives. It does not require destruction of legally retained network logs, transactions, complaints and safety-event evidence where retention is necessary. Explain separately “delete relationship memory”, “remove chat display”, “close the account” and “legally retained record”.

When a game sends player dialogue to a foundation-model provider, it should identify whether the provider is a processor acting on instructions, a joint decision-maker or an independent user. The contract should address processing-only use, no-return training, human access, subprocessors, retention, cross-border transfer, incidents, audits and deletion proof. If a provider actually uses dialogue for its own model improvement, the legal analysis must follow that real purpose rather than the contract label.

The Measures do not ban every use of interaction data for training. They prohibit sensitive personal information contained in user interaction data from being used for model training unless another law applies or the user gives separate consent. Mental health, medical, financial, precise-location, biometric and under-fourteen information can be sensitive. Separate consent should not be hidden in general terms or forced by making the core game unusable. Defaulting to no training and separately authorising carefully scoped samples is the safer design.

7. Assessment, filings, stores and responsibility

The Measures identify five safety-assessment triggers: launching a service or adding a relevant function; a major change caused by new technology; 1 million registered users or 100,000 monthly active users; a risk that may affect national security or the public interest; and other prescribed or notified circumstances. The report covers safety controls, training data, extreme scenarios, user scale and age, minors and elderly users, complaints and remediation. Launch itself is an independent trigger; the size threshold is not a small-team exemption.

The text does not say that every new character needs a separate assessment. A cosmetic or narrative change with the same model, memory, topics, age access and business model may not be a major change. The need to re-assess rises sharply when a task assistant gains real-life topics, cross-session memory, voice or emotion recognition, minor access, intimacy payment, a stronger model or open public UGC. Maintain a versioned change ledger and record why a change was or was not material.

Article 26 connects anthropomorphic services to algorithm recommendation filing, changes and cancellation, with annual verification. A game that directly provides generative AI must separately consider generative-AI service filing, model registration, algorithm filing and the anthropomorphic-service assessment. Using a registered foundation model does not finish the game’s own service review.

App stores must verify the relevant assessment and filing and can refuse listing, warn, suspend or remove a non-compliant app. The store check is not a liability shield for the provider, and a developer promise is not a substitute for a store’s required verification. Developers, publishers and operators should allocate who controls the role, data, payment, crisis escalation, filings, store materials and shutdown. A contract saying “the publisher handles local compliance” is incomplete where both sides determine the feature.

8. What recent product adjustments do—and do not—prove

Public reporting in July 2026 described adjustments to human-like and user-created agents on major Chinese AI platforms around the effective date of the Measures. The reports can support three limited facts: some platforms changed or ended certain agent functions; names of some affected functions referred expressly to anthropomorphic interaction or user creation; and some notices provided export or later data handling. They do not prove that an authority directly ordered a shutdown or that the platforms admitted a violation. [10][11]

An open user-created agent platform faces a different risk surface from a constrained game NPC: topics, relationship design, age access, memory and review capacity differ. A shutdown report should therefore not be used to infer that every game AI NPC is unlawful. It is better read as evidence that platforms are separating controllable business agents from open relationship features and preserving user-data transition paths.

9. Comparative law: similar signals, different structures

The EU AI Act Article 50 primarily addresses transparency. An interactive AI should disclose its AI identity, and providers have machine-readable marking duties for synthetic outputs. Emotion-recognition deployments have separate notice requirements. The general application date is 2 August 2026. [12] It does not create China’s prohibition on virtual intimacy for minors, two-hour reminder, safety assessment or algorithm filing.

California SB 243 defines a companion chatbot around a natural-language interface, adaptive human-like responses, social needs and a relationship maintained over multiple interactions. It creates a narrow exclusion for a game bot whose responses are only about the game, do not discuss mental health, self-harm or explicit sexual content, and do not maintain non-game dialogue. This is a functional boundary, not a general game exemption. The statute also addresses AI disclosure, minor safety, crisis protocols and a private remedy. [13][14]

New York General Business Law Article 47 focuses on systems that remember interactions or preferences, proactively ask emotional questions outside the immediate response and continually discuss personal matters. It excludes customer service, productivity, research and enterprise assistance. Operators must maintain crisis protocols and give identity and time reminders. [15][16]

The US Federal Trade Commission’s 2025 Section 6(b) inquiry into companion chatbots studied monetisation, child and adolescent effects, development and review, disclosure, age limits and personal-information use. It was a study request, not a finding of violation. [17] The common lesson is AI identity, crisis handling and minor protection; the Chinese product team should not copy a US hour count as a Chinese safe harbour.

10. A four-layer decision framework

Use four layers of evidence:

1. Technology and public provision: is AI actually used, and is the service offered to the mainland public?
2. Anthropomorphic role: does the system present a natural-person-like identity and communication style?
3. Continuing relationship: does it use memory, long conversations, a recurring entry, proactive recall or relationship stages?
4. Emotional service: is care, companionship, support, intimacy or dependence part of the service or commercial objective?

On this framework:

• fixed-script characters; narrow task, combat and rule assistants; and non-continuing customer service or Q&A are generally outside the core object;
• stable-persona, no-memory open dialogue; task teammates that occasionally comfort; limited real-life chat; and narrative emotional stages require factual review; and
• characters with cross-session memories of real-life experience and proactive care, paid romance or family replacement, dependence mechanisms, and open public platforms for emotional personalities are likely in the core zone.

Evidence matters more than naming. If a company says a feature is only a task NPC, it should retain topic boundaries, prompts, memory fields, proactive-message rules, age access, marketing, payment mechanics, tests and exit logic. If an update crosses the boundary, re-assessment should follow.

11. Conclusions and limits

The central question is not “does the character use a large model?” It is “is the product operating a continuing human-AI emotional relationship for the mainland public?” Fixed scripts, task intelligence, stable personas, memory and virtual companions are legally distinguishable. The rules should preserve those distinctions.

Where the Measures apply, obligations reach the dialogue interface, session timer, real-name and age system, crisis escalation, memory and data pipeline, release version and store submission. The largest operational consequences may be feature restriction, registration suspension, service termination and delisting rather than the fine alone.

Five limits remain material: official guidance has not fully defined continuity, emotional interaction or major change; there is no public game-specific enforcement sample; the mainland-public connection for global servers and foreign platforms remains fact-sensitive; emergency-contact collection must be balanced with minimisation and false alarms; and the need to re-file after a model, character or UGC change will depend on provincial practice.

References

[1] Cyberspace Administration of China, Q&A on the Interim Measures for the Administration of Anthropomorphic Interaction Services, source

[2] Cyberspace Administration of China et al., Interim Measures for the Administration of Anthropomorphic Interaction Services, source

[3] Cyberspace Administration of China et al., Interim Measures for the Administration of Generative Artificial Intelligence Services, source

[4] Cyberspace Administration of China et al., Provisions on the Administration of Algorithmic Recommendations in Internet-based Information Services, source

[5] Cyberspace Administration of China et al., Provisions on the Administration of Deep Synthesis in Internet-based Information Services, source

[6] Cyberspace Administration of China et al., Measures for Labelling AI-Generated and Synthetic Content, source

[7] Cyberspace Administration of China, Draft Interim Measures for the Administration of Anthropomorphic Interaction Services, source

[8] Cyberspace Administration of China, “Shanghai’s first-stage campaign against AI application disorder,” source

[9] National Press and Publication Administration, “Notice on Further Strictly Preventing Minors from Becoming Addicted to Online Games,” source

[10] IT Home, report on Qwen’s human-like and user-created agent adjustments, source

[11] Yangcheng Evening News, report on adjustments to Doubao and Qwen agents, source

[12] Regulation (EU) 2024/1689, Article 50, source

[13] California SB 243, §22601, source

[14] California SB 243, §§22602–22605, source

[15] New York General Business Law §1700, source

[16] New York General Business Law §§1701–1703, source; source; source

[17] Federal Trade Commission, “FTC Launches Inquiry into AI Chatbots Acting as Companions,” 2025-09-11, source