China's Practice and Institutional Improvement of Privacy Protection for Embodied AI: From Data Compliance to Contextual Compliance

1. Background

Embodied artificial intelligence refers to physical intelligent agents capable of perception, mobility, and autonomous decision-making. These agents learn and evolve through real-time interaction with the physical world. Embodied AI products, represented most visibly by humanoid robots, are moving from laboratories toward large-scale commercial deployment and show considerable potential in home companionship, industrial collaboration, medical care, and public services. According to existing projections, the scale of China’s embodied AI industry is expected to reach 400 billion yuan by 2030 and to surpass one trillion yuan by 2035.

Compared with conventional AI, embodied AI collects data comprehensively through multimodal sensors, including visual, auditory, tactile, and light detection and ranging sensors. Its mobile physical form can actively enter private physical spaces, while algorithmic opacity may generate emergent autonomous behavior. The associated privacy risks are continuous, imperceptible, and context-specific. They extend beyond conventional static privacy boundaries and reach into the most private physical spaces and everyday settings of natural persons, exposing the limitations of the traditional data compliance paradigm centered on notice and consent (Li & Liang 2026).

No unified global consensus has yet emerged on privacy protection for embodied AI. China is among the countries in which the embodied AI industry is advancing most rapidly. While promoting industrial innovation, it is also exploring a governance approach that coordinates high-quality development with a high level of security. Its current framework rests on three foundational statutes, the Cybersecurity Law, the Data Security Law (DSL), and the Personal Information Protection Law (PIPL), supplemented by the Regulations on Network Data Security Management and regulatory rules governing algorithms and generative AI. Local legislation and a national system of technical standards provide additional context-specific support, allowing the framework to respond dynamically to privacy concerns arising in different embodied AI applications. Against this background, a systematic examination of China’s legal practice and the construction of a governance framework suited to the development of embodied AI carry substantial theoretical and practical significance.

This article examines how embodied AI challenges the existing privacy-protection system. It reviews China’s legal, regulatory, and technical standards at the normative level, showing how they coordinate to address gaps and build a dynamic, context-sensitive governance framework. The article then identifies the strengths and limitations of China’s current approach and advocates a governance model grounded in contextual integrity theory, with improvements proposed across legislation, regulation, and standards. Balancing innovation incentives with privacy protection, this study offers China’s institutional experience as a constructive reference for global AI governance.

2. Governance Challenges Posed by the Distinctive Privacy Risks of Embodied AI

2.1. Technical Features That Distinguish Embodied AI from Conventional AI

As early as 1950, Turing foreshadowed the possibility of embodied intelligence, envisaging machines that could perceive their environment, reason, make decisions, and act in ways comparable to human beings. (Turing, 1950) This vision represents the ultimate form of AI development. Embodied AI therefore differs in important respects from conventional AI applications that exist primarily in virtual environments, creating new risks for personal privacy under traditional data protection frameworks.

In terms of perception, conventional AI generally obtains personal data through active user input. Data collection is intermittent and perceptible to the user, as illustrated by large language models such as GPT, Claude, and Gemini. Embodied AI, by contrast, uses multimodal sensors to collect diverse forms of data from its surroundings continuously and often without the affected person’s awareness. Article 1035 of the Civil Code of the People’s Republic of China‌ requires the personal information of a natural person shall be processed under the principles of lawfulness, justification and necessity, shall not be excessively processed. Article 14(1) of PIPL further provides that consent to personal information processing must be voluntary, explicit, and fully informed. Together, these provisions establish notice and consent as a basic principle governing data collection and processing in China. Continuous and imperceptible collection weakens the substantive force of that principle, which is grounded in the individual’s right to informational self-determination. Users cannot remain constantly attentive to an embodied AI system’s collection practices and may have no genuine understanding of when, where, or what data it gathers. The system may also collect highly sensitive personal information, including biometric information, health information, and details of everyday habits, without the user’s knowledge.

In terms of decision-making, conventional AI generally relies on predefined algorithms and explicit instructions, and its data processing activities ordinarily remain within an initially prescribed scope. Embodied AI can learn and evolve autonomously. Through real-time physical interaction with its environment and the feedback generated in that process, it undertakes dynamic and autonomous deep learning and control-related decision-making. This emergent autonomy makes the purposes of data processing unclear and indeterminate (Xu et al. 2025). A home companion robot, for example, may initially be designed to provide everyday services. Over the course of prolonged interaction, however, it may learn private information about the user’s health or emotional changes and use that information for other commercial purposes that were not disclosed in advance.

In terms of its mode of existence, conventional AI processes data mainly within the virtual environment of the internet, where privacy protection focuses on preventing the unlawful acquisition and use of personal information online. Embodied AI has a physical presence. It can enter a natural person’s private space and engage in close physical interaction, thereby crossing the boundaries of conventional online privacy. Privacy in embodied AI applications consequently encompasses personal information protection as well as a natural person’s rights to spatial tranquility, decisional autonomy, and freedom from imperceptible intrusion in physical space (Li & Chen 2025).

Embodied AI thus enters living spaces and networks of social relations through hardware devices. Its privacy effects often develop gradually through sustained human interaction and cannot readily be reduced to a single act of data collection, a particular use of data, or one stated processing purpose. This feature presents a serious challenge to traditional legal and regulatory systems.

2.2. Types of Privacy Risk in Typical Application Contexts

The perceptual, decisional, and physical characteristics of embodied AI make it necessary to assess privacy risks within the specific contexts in which these systems are used.

2.2.1. Domestic Settings

The inviolability of residence constitutes a core personality interest jointly established by constitutional and civil law, and the domestic space enjoys absolute seclusion under the traditional privacy framework (Xu 2023). Embodied AI devices, including domestic service robots and intelligent interactive terminals, can continuously collect data on spatial layouts, movement trajectories, audiovisual conversations, and environmental characteristics even when no person is actively interacting with them. Such data can reconstruct deeply private aspects of household life, including family members’ health conditions, distribution of property, patterns of intimate relationships, and the developmental trajectories of minors. Behavioral patterns may also support inferences about living habits, consumer preferences, and emotional states.

Temporary occupants such as visitors are unable to participate meaningfully in the notice-and-consent mechanism and have no effective channel through which to control data processing. In many circumstances, they are entirely unaware that they fall within a sensor’s collection range. More alarmingly, embodied AI deployed in domestic settings possesses affective interaction capabilities. By continuously collecting emotional data and conversational preferences, these systems can build user profiles that algorithmic recommendation systems may carry into consumption decisions, choices concerning household affairs, and even value judgments. The resulting risk of affective manipulation exceeds the conventional domain of informational privacy and directly interferes with a civil subject’s freedom to form intentions.

2.2.2. Public Settings

In public settings, the principal privacy risks arise from imperceptible surveillance and infringements affecting an indeterminate number of persons. Individuals in public spaces retain a reasonable expectation that they will not be specifically identified or continuously tracked. Facial recognition access systems in residential compounds, sensing devices in commercial premises, external sensors on intelligent connected vehicles, and public service robots can nonetheless collect faces, movement trajectories, behavioral characteristics, and other sensitive biometric information continuously, without the individual’s knowledge or any active interaction (Chen 2022).

Unlike the risks experienced by identifiable persons in a home, public-space collection affects an indeterminate population and involves multiple categories of data subjects. In 2026, for example, the Liangjiang New Area People’s Procuratorate in Chongqing found in the course of performing its duties that a facial recognition system operated by an enterprise in a residential compound stored more than 100,000 facial records. The database included information concerning children under the age of 14, although the separate consent of their guardians had not been obtained.( The Supreme People’s Procuratorate of the People’s Republic of China 2026) The risks associated with processing on this scale extend beyond an infringement of one person’s private rights and implicate the public interest of an indeterminate group. Because the cost of enforcing individual rights against large-scale processing is extremely high, procuratorial public interest litigation is needed as a governance mechanism.

Data collection in public settings is also frequently tied to access to public services or eligibility to complete ordinary transactions. An individual who refuses facial collection may be unable to enter a residential compound, use a public service, or complete a basic purchase. These conditions undermine the voluntary character of notice and consent.

2.2.3. Industrial Settings

In industrial settings, the central risks concern the disclosure of trade secrets and production safety, together with intertwined interests in property, public safety, and workers’ personality rights. Privacy risks in homes and public spaces generally involve infringements of personality rights or resulting financial harm. Industrial embodied systems, including industrial robots, inspection robots, and intelligent warehousing equipment, are directly embedded in production processes. Tampering with sensor data or malicious interference with automated decisions can cause equipment to lose control, disrupt a production line, trigger an industrial safety accident, and result in casualties or substantial property loss.

Governance of embodied AI in industrial settings therefore cannot rely solely on personal information protection rules and must be coordinated with the regulatory systems governing product safety and workplace safety. Industrial applications also raise concerns for workers’ personality rights. Data collected on workers’ operational behavior, biometric characteristics, and movement at work may enable algorithmic surveillance or excessive labor demands. Within the hierarchy of protected legal interests in this setting, these concerns remain subordinate to production safety and enterprise data security, and their governance must also be coordinated with labor protection rules rather than addressed exclusively through general personal information protection requirements.

3. Legal Framework and Practice for Privacy Protection in Embodied AI in China

The global embodied AI industry is entering the initial phase of large-scale deployment. Based on their respective legal traditions and industrial positioning, major countries and regions have developed three differentiated privacy governance pathways, yet a mature and unified institutional framework has yet to emerge.

The EU adopts a binding, risk-based model centered on the Artificial Intelligence Act and the General Data Protection Regulation, classifying most caregiving and domestic embodied AI agents as high-risk and mandating pre-market compliance and impact assessments. Complementary instruments, including the Data Act, Cyber Resilience Act, and revised Product Liability Directive, extend governance to data access, product security, and liability. While the EU framework excels in structural coherence and cross-border coordination, its high compliance costs disproportionately burden small innovators, and its underlying online-data logic remains ill-suited to embodied AI’s spatial mobility and continuous physical-world data collection.

The United States relies on decentralized state-level regulation and industry self-regulation. This approach offers regulatory flexibility and gives individual enforcement actions a strong deterrent effect. In 2023, the Federal Trade Commission and the Department of Justice brought an enforcement action concerning Amazon Alexa’s processing of children’s voice data and required the deletion of children’s data as well as certain voice and geolocation data (United States v. Amazon.com, Inc. 2023). On the same day, the Federal Trade Commission commenced proceedings against Ring, alleging that employees and contractors had improperly accessed private videos and that the company had failed to implement basic security safeguards (Federal Trade Commission v. Ring LLC 2023). These actions operate alongside the Health Insurance Portability and Accountability Act, which requires administrative, physical, and technical safeguards for electronic health information; the Children’s Online Privacy Protection Act, which requires verifiable parental consent for information concerning children under the age of 13; the California Consumer Privacy Act; and the Illinois Biometric Information Privacy Act. Taken together, these rules allow regulators to address privacy risks in particular embodied AI contexts with considerable precision. Their fragmented sources and complex boundaries nevertheless require enterprises to navigate federal sectoral law, state law, and the Federal Trade Commission’s general enforcement authority simultaneously.

Japan and the Republic of Korea both pursue flexible governance based on industrial coordination. The intensity of their regulatory approaches falls between the European Union’s stringent legislation and the United States’ reliance on industry self-regulation. Japan’s Act on the Protection of Personal Information establishes foundational rules on consent, security management, and breach notification. Its Act on the Promotion of Research, Development, and Utilization of Artificial Intelligence-Related Technologies centers on voluntary compliance, while the Personal Information Protection Commission primarily uses administrative guidance and imposes comparatively limited sanctions.

The Republic of Korea’s AI Basic Act, which took effect in 2026, applies tiered regulation to high-risk AI, incorporates privacy and security into product certification, and uses local pilot programs to implement context-specific rules. Its regulatory requirements are somewhat more binding than those of Japan. Both countries prioritize industrial development and display a distinct coordination of hard and soft law, although each provides insufficient specialized regulation of risks unique to embodied AI, including privacy in physical space.

China currently extends existing rules, particularly PIPL, to the governance of privacy in embodied AI and has not enacted dedicated legislation. The following analysis examines the operation of this framework through four dimensions: legislative supply, judicial determination, regulatory enforcement, and contextual application.

3.1. Legal Foundations for Privacy Protection in Embodied AI

3.1.1. Support from the Data Security Legal Framework

The rapid development of China’s embodied AI industry has prompted extensive debate about its legal risks. The prevailing view in Chinese scholarship and practice holds that the application of embodied AI is already subject to law. The Civil Code, PIPL, DSL, Cybersecurity Law, and Regulations on Network Data Security Management collectively provide the legal foundation for protecting privacy in embodied AI.

The Civil Code establishes the basic framework for privacy protection in China. Article 1032 declares that a natural person enjoys the right of privacy and prohibits infringement by spying, intrusion, divulgence, public disclosure, or any other means. It defines privacy as the tranquility of the private life of a natural person, together with the private space, private activities, and private information that he is unwilling to be known to others. The protected legal interests therefore extend beyond mere data leakage incidents. They encompass continuous intrusions by embodied AI into private physical space, interference with the tranquility of private life, and impairment of personal dignity. Article 1032 thus supplies a civil-law baseline for regulating information collection in embodied AI settings.

Building on this foundation, PIPL delineates the legal boundaries for processing sensitive personal information, including biometric information, movement trajectories, health information, and financial account information, and takes the protection of individual interests as its point of departure (He 2024). Guided by the protection of individual interests, PIPL permits the processing of sensitive personal information only where there is a specific purpose and a compelling necessity, coupled with the adoption of rigorous protective measures. Furthermore, the processing of biometric information requires the separate consent of the individual. These provisions establish clear legal requirements for the protection of biometric data in embodied AI contexts.

DSL, approaching the matter from the perspective of "data as a security object," brings the security of data flows involving embodied AI—including training datasets, industrial operational data, and sectoral data—within the regulatory perimeter through its institutions of data classification and tiered protection, catalogs of important data, and data security risk assessment. Article 20 of the Cybersecurity Law incorporates AI security and development into the cybersecurity legal system at the statutory level, thereby providing an overarching statutory gateway for training data compliance, security of computing infrastructure, and algorithmic security assessment and regulatory interfaces in the embodied AI domain. As an administrative regulation, the Regulations on Network Data Security Management further consolidate the responsibilities of network data processors and impose more operational compliance duties where important data or personal information is processed on a comparatively large scale.

3.1.2. Context-Specific Governance Tools under Sectoral Regulatory Rules

China’s sectoral regulatory system for AI remains applicable to the novel privacy risks generated by embodied AI. It is oriented toward specific applications and organized around control throughout the data lifecycle. The system reaches beyond static data compliance requirements by embedding privacy obligations in concrete operational settings, providing a foundation for contextual privacy governance adapted to embodied AI.

At the level of privacy governance for generated content, the Interim Measures for the Management of Generative Artificial Intelligence Services and the Provisions on the Administration of Deep Synthesis of Internet Information Services form a central line of protection. For embodied AI systems equipped with large models, these rules require the cleaning and preprocessing of training data to comply with national security and ethical standards. The Measures for Labeling Artificial Intelligence-Generated and Synthetic Content further establish dual obligations to apply explicit and implicit labels. When an embodied AI system conducts a voice conversation, generates an image, or presents a virtual persona, its output must therefore remain identifiable and traceable. These requirements address identity appropriation and privacy leakage caused by deepfakes or deceptive interaction at their source.

At the level of privacy protection in algorithmic decision-making, the Provisions on the Administration of Algorithmic Recommendations for Internet Information Services perform an algorithmic auditing function (Zhang 2024). Task planning and autonomous decision-making by embodied AI depend substantially on algorithmic recommendation and automated decision-making. The provisions require the establishment of review mechanisms for algorithmic models and mechanisms, together with science and technology ethics review. These requirements seek to penetrate the algorithmic black box and ensure that physical actions such as obstacle avoidance, grasping, and path planning do not result in discriminatory treatment of minors, older persons, or other particular groups, or cause damage to the physical environment because of algorithmic bias or defects. In this way, they provide indirect protection for personal privacy.

Specialized regulation of anthropomorphic interaction is especially important. The Interim Measures for the Administration of Anthropomorphic AI Interactive Services identify the risks of psychological dependence and emotional manipulation associated with embodied AI. Directed specifically at services that simulate personality traits, the measures grant users control rights that include the rights to delete and copy their data and expressly prohibit the use of anthropomorphic interaction to infringe human dignity. They directly constrain robots used for home care and emotional companionship and provide an enforcement basis for preventing embodied AI from inducing excessive disclosure of private information or exploiting emotional dependence to exercise psychological control.

3.1.3. Technical Oversight through Industrial Policies and Standards

China has made technical standards an implementation instrument for privacy governance in embodied AI. It is gradually developing a system of technical oversight that covers the full lifecycle of embodied AI and translates the general requirements of statutes and departmental rules into quantifiable and verifiable engineering criteria. In February 2026, the Standardization Administration of China and the Ministry of Industry and Information Technology jointly issued the Humanoid Robots and Embodied AI Standard System (2026 Edition). As the world’s first national-level specialized standard system for embodied AI, its safety and privacy component addresses privacy risks in embodied AI applications. It specifies technical requirements for boundary controls on multimodal sensing, thresholds for biometric information collection, privacy protection in edge computing, and user-perceptible indications of privacy status. These requirements anchor legal principles such as notice and consent under PIPL in concrete technical parameters.

Industrial policy also incorporates privacy requirements throughout the lifecycle of embodied AI. The Guiding Opinions on the Innovative Development of Humanoid Robots, issued by the Ministry of Industry and Information Technology, calls for an effective privacy-by-design governance mechanism while integrating AI, advanced manufacturing, new materials, and other technologies to promote industrial development. It thereby makes privacy compliance a precondition of product research and development. The Guidelines for the Construction of a Comprehensive Standardization System for the National Artificial Intelligence Industry identify embodied AI as a key area. They provide for supporting privacy standards covering multimodal active interaction, autonomous behavioral learning, simulation, and other core technical stages, and clarify requirements for data desensitization during interaction and privacy impact assessment during autonomous learning.

Embodied AI action plans in Beijing, Shenzhen, and other localities further implement privacy standards in specific settings. Beijing is constructing an embodied AI testing and validation facility that includes a dedicated privacy compliance testing area to simulate the privacy requirements applicable to home care, community services, and other uses. Shenzhen requires all embodied devices deployed in its pilot programs for open application settings to comply with technical standards on local storage, traceability of conduct, and one-click sensor shutdown. These initiatives provide practical examples of the industrial implementation of privacy requirements.

3.2. Representative Judicial Cases Involving AI-Related Privacy Infringement

China has not yet produced a legally effective judgment directly involving an embodied AI system as the alleged infringing actor. Cases concerning personality rights and personal information protection in the use of AI deep synthesis and biometric information have, however, generated adjudicatory rules whose logic can extend directly to embodied AI. These rules establish clear judicial boundaries for privacy protection under multimodal sensing.

3.2.1. Integrated Protection for the Digital Extension of Personality Attributes

Where an enterprise fixes a natural person’s portrait, voice, or behavioral patterns as the interactive persona of an embodied AI system, it must obtain specific authorization covering the relevant personality rights even if the persona is carried by a physical robot and used to provide everyday services. The robot’s operator bears primary responsibility for the personified content of its interactions and cannot avoid infringement liability by characterizing that content as autonomously generated by an algorithm. He v. Shanghai Artificial Intelligence Technology Co. (2021) was China’s first legally effective judgment systematically to establish rules protecting personality rights in AI-generated virtual personas. [1]The case reflects an expansive judicial approach under which technological development does not diminish the protection afforded to personality rights. First, the scope of protection extends with changes in technological form. A digital virtual persona or the appearance of a physical robot falls within the law of personality rights whenever it permits identification of a particular natural person. Second, providers of algorithmic services owe a duty of care exceeding that of ordinary internet service providers and must actively review and control the risk of infringing representations of human personality generated through their platforms.

3.2.2. Identifiability as the Standard for Protecting Voice-Related Interests

An embodied robot that collects or replicates a natural person’s voice for routine interaction or model optimization must obtain separate and explicit consent at the level of personality rights. Training on publicly available voice recordings without authorization to use the speaker’s personality attributes may likewise infringe the legally protected interest in that person’s voice. Yin v. Beijing Intelligent Technology Co. (2024) was China’s first case concerning infringement of the right to voice through AI. [2]The plaintiff, a professional voice actor, alleged that the defendants had used her publicly available recordings without authorization to train an AI voice model, generated a highly similar synthetic voice product, and sold it commercially through a platform. The judgment made identifiability the central standard for protecting voice-related interests. When an AI-generated voice enables the ordinary public to identify a particular natural person, it falls within the Civil Code’s protection of voice-related personality interests. The judgment filled a gap in the protection of biological attributes of personality in AI synthesis and confirmed that technical processing does not sever those attributes from the person. This rule provides a clear compliance boundary for embodied robots equipped with voice interaction and voiceprint recognition.

3.2.3. Strict Regulation of the Misuse of Biometric Information

Biometric information is unique and cannot be altered, and the harm caused by its disclosure is irreversible. Its collection, use, and processing are therefore subject to substantially stricter regulation than ordinary personal information. A person who processes another individual’s biometric information in bulk for the purpose of creating and distributing deepfakes for profit may incur criminal liability, civil infringement liability, and liability to compensate for harm to the public interest. Embodied AI routinely collects facial images, voiceprints, and other biometric data through visual and auditory sensors and consequently falls within a high-risk regulatory category. An enterprise that fails to discharge its compliance obligations or unlawfully stores or misuses biometric information may incur civil liability. Serious circumstances may also create criminal exposure and liability through public interest litigation. In the criminal case and incidental civil public interest action against Yu for using AI face-swapping technology to infringe citizens’ personal information, Yu illegally obtained facial information, used deep synthesis technology to create and disseminate obscene videos online, provided customized AI face-swapping services, and sold related software for profit. [3]Chinese courts treated facial information as highly sensitive biometric information, defined the criminal boundary of its misuse, and used public interest litigation to provide residual protection for the informational interests of an indeterminate number of people. The resulting standard is one of comparatively strict protection.

3.3. Contextual Regulatory Enforcement Involving Multiple Actors

Regulatory authorities at all levels consistently seek to balance development with security, while local governments serve as important sites for industrial implementation and regulatory experimentation. Hangzhou is a major center of China’s embodied AI industry. Building on an industrial base of more than 700 robotics enterprises and output valued at over RMB 100 billion, the city provides multidimensional policy support through fiscal, financial, technological, and land-use measures. It focuses on cultivating demonstration applications in home, medical, and industrial settings and explores context-specific privacy rules as deployment at scale proceeds, allowing industrial development and risk governance to advance together (Hangzhou Municipal Bureau of Economy and Information Technology & Hangzhou Municipal Bureau of Finance 2026).

China has also developed a system of regulatory enforcement that involves multiple actors and covers the full operational process. Administrative regulators and judicial authorities work in coordination. Cyberspace authorities use administrative compliance consultations and targeted inspections, while procuratorial authorities issue procuratorial recommendations to address unlawful facial data collection in scenic areas, supermarkets, and other public places. Enterprises can be required to delete unlawfully collected information promptly and complete technical upgrades, enabling precise governance of frequently occurring risks.

Regulatory intervention is also moving upstream. In autonomous driving, healthcare, and other fields, sandbox security-service platforms and compliance testing guidelines encourage enterprises to align security and privacy technologies during research and development and shift data compliance from ex post remediation to ex ante incorporation. Industry self-governance supplies an additional component. The Artificial Intelligence Industry Alliance of China has issued a Self-Regulatory Convention on Privacy Protection for Embodied AI, calling on enterprises to comply with laws and regulations, respect user privacy, and adopt the technical and managerial measures necessary to protect personal information. Several leading embodied AI enterprises have signed the convention and committed to stronger self-regulation and the maintenance of sound market order.

This flexible model of co-governance is a localized application of coordinated hard and soft law in an emerging technological field. Its principal advantage lies in its capacity to respond to the delayed operation of written law and to accommodate the dynamic and context-specific development of embodied AI at a lower institutional cost of experimentation (Li 2025). Regulatory sandboxes provide space to test compliance rules across settings of differing sensitivity. Administrative guidance and compliance consultations reduce the inhibitory effect that one-size-fits-all binding regulation may have on technological innovation.

The current flexible governance system is generally suited to the early cultivation of the embodied AI industry and provides an institutional basis for implementing contextual compliance. As embodied AI is adopted at scale, further work will be needed to connect soft-law norms with mandatory legal rules and to clarify the legal effect and procedural boundaries of flexible regulatory measures.

4. Improving China’s Privacy Protection Regime for Embodied AI

China attaches importance to preventing the security risks of embodied AI while promoting industrial innovation. This governance approach protects user privacy and leaves sufficient room for industrial development. Although China has made progress in protecting privacy in embodied AI, several elements of the existing regime remain open to further improvement.

4.1. Contextual Compliance as the Basic Governance Approach

4.1.1. The Perspective of Contextual Integrity

China has developed an initial legal governance framework for privacy in embodied AI that encompasses legislation, standards, technology, and regulation. Its underlying logic nevertheless remains rooted in a static model of data protection centered on notice and consent and grounded in the individual’s right to informational self-determination. That model arose in online internet settings. It now confronts the imperceptible and continuous collection of data by embodied AI, the emergent and opaque character of algorithmic decisions, and the mobility and intrusiveness of physical intelligent agents.

Helen Nissenbaum’s theory of contextual integrity provides an important theoretical perspective for contextual compliance. In her account, the privacy implications of information do not turn exclusively on the sensitivity of the information itself. The decisive question is whether information flows comply with the norms that govern information in a particular social context (Nissenbaum 2004). When the social status of embodied AI remains incomplete and the relevant context is uncertain, the corresponding legal relationships also remain uncertain (Balkin 2015). People hold different reasonable expectations regarding information flows in different contexts, and privacy is preserved when those flows conform to the relevant expectations. The uncertainty surrounding legal relationships involving embodied AI arises fundamentally from its cross-context capacity as a mobile physical actor. When an embodied AI system moves from a private home into public space, or from a production setting into a service setting, the norms governing information flows in the original context cease to apply. A legal regime that imposes fixed and undifferentiated duties on data processing across all settings will inevitably produce a mismatch between law and fact.

The theory of contextual integrity exhibits a notable degree of congruence with PIPL. Article 5 of PIPL stipulates that "personal information shall be processed under the principles of legality, legitimacy, necessity, and good faith, and no organization or individual may process personal information by misleading, defrauding, coercing, or through any other such means." Article 6 provides that "the processing of personal information shall have a clear and reasonable purpose, be directly relevant to the processing purpose, and be conducted in a manner that has the minimum impact on individuals’ rights and interests." These provisions embody the philosophy of contextual governance, emphasizing that compliance obligations should be determined by reference to the specific processing purpose and the actual impact on individuals’ rights and interests.

Unlike static data compliance, which centers on fixed rules irrespective of context, contextual compliance directs its attention to the specific scenario in which data processing occurs and to the actual effect on the rights and interests of natural persons. It requires enterprises to formulate differentiated compliance strategies tailored to the risk level and reasonable expectations inherent in each scenario. Contextual compliance underscores ex ante contextual privacy impact assessments, intra-process dynamic monitoring, and ex post accountability, thereby offering a governance model far better adapted to the distinctive developmental characteristics of embodied AI.

4.1.2. Reconstructing the Traditional Paradigm through Contextual Compliance

The first implication of contextual compliance is reduced dependence on notice and consent doctrine. The traditional paradigm assumes that a rational user can provide informed consent. In embodied AI applications, users may be unable to make an effective decision during complex, continuous, and imperceptible interaction. An agent’s anthropomorphic appearance and emotional inducement may also prompt authorization in a non-rational state that is not genuinely voluntary. Legal protection should therefore give greater weight to whether data flows in the relevant setting conform to the reasonable expectations of the public than to the formal presence of consent. Consent nevertheless remains part of the framework when its relative role is reduced (Tschider 2021). The reasonableness of information processing should be assessed by reference to contextual elements such as the identity of the data subject, the purpose of processing, and the type of information. From the perspective of a reasonable person, the inquiry asks whether processing exceeds a reasonable expectation regarding the flow of information (Liu 2021).

A second implication is the dynamic calibration of purpose limitation. The emergent character of embodied AI allows processing purposes to expand and generate derivative uses during operation, depriving a static rule based on one-time authorization of much of its practical effect. Contextual theory permits dynamic recalibration because the norms governing information flows must change with the context. The legitimacy of processing depends on continuing compliance with the information-flow norms of the initial setting rather than on a single purpose identified in advance (Wang 2022). When an embodied AI system moves from a home setting into a medical setting, or when autonomous learning generates a new use for data, the context must be reassessed and the applicable distributional rules updated.

A third implication concerns flexible application of the principle of minimum necessity. Article 6 of PIPL stipulates that the collection of personal information shall be limited to the minimum scope necessary for achieving the processing purpose, and prohibits excessive collection. It further provides that processing shall be conducted in a manner that has the minimum impact on individuals’ rights and interests. Previous applications of the principle attempted to define the boundary of collection through quantitative standards, an approach that is difficult to sustain in multimodal and cross-context data flows. Consequently, the necessity of data collection must be gauged by reference to the functional logic of the specific scenario and the reasonable expectations of users. For instance, when a family companion robot is executing security patrol tasks, capturing visual data within bedrooms may be deemed necessary. By contrast, if the same robot simultaneously engages in user behavior analysis for commercial product recommendations, that very act of data collection would lose its legitimate justification in a commercial context.

4.2. Legislative Improvement

4.2.1. Establishing the Basic Rules of Contextual Compliance

Because the embodied AI industry remains in a period of rapid iteration, institutional design must combine stability with flexibility. In the short term, priority should be given to amending the PIPL, DSL, and other existing statutes. Special provisions should address data processing by embodied AI and the protection of spatial privacy, while contextual assessment should form part of the legal standard for determining whether personal information processing is lawful. These amendments would address gaps in the application of existing rules. The State Council should also formulate Regulations on the Development and Security Management of the Embodied AI Industry. An administrative regulation could establish a basic framework for industrial oversight, risk prevention and control, and contextual classification and respond quickly to the practical needs of industrial development. Once industrial forms and risk characteristics become more stable, the Standing Committee of the National People’s Congress could enact a dedicated Artificial Intelligence Law of the People’s Republic of China (Zhang 2024). Such a law would define systematically the legal status, foundational principles, and core institutions governing the development and use of embodied AI.

At every stage of this legislative process, contextual compliance should be established as a core principle of privacy protection for embodied AI. Contextual integrity should provide the theoretical foundation for the design of rules, supporting a differentiated system that classifies contexts and allocates responsibility according to risk so that the intensity of privacy protection corresponds precisely to the risk level of each setting.

4.2.2. Clarifying the Legal Status and Statutory Definition of Embodied AI

Legislation should first clarify two foundational rules and thereby resolve threshold disputes concerning the application of the legal framework. The legal nature of an embodied intelligent agent should be defined without recognizing independent legal personality. Embodied agents should instead be classified as AI products and systems capable of autonomous perception and decision-making. The legal consequences of their operation should be borne in accordance with law by the relevant producer, algorithm developer, service operator, and user. This allocation would prevent autonomous decision-making from diluting responsible parties or creating a responsibility vacuum (Mei 2025). Legislation should also adopt a uniform statutory definition of embodied AI and classify systems into different risk levels according to perceptual capacity, degree of autonomy, and application context, thereby supplying a uniform legal baseline for subsequent classified and tiered regulation. The statutory criteria for classifying contextual risk should use spatial sensitivity and functional attributes as their central dimensions. Application contexts could then be divided into highly sensitive private settings, ordinary living settings, public service settings, and industrial production settings, establishing clear legal boundaries for the application of contextual compliance rules.

4.2.3. Refining Personal Information Protection Rules

The principle of data minimization should be further specified. An embodied AI enterprise should collect only the minimum data necessary to perform a particular function and should not collect data unrelated to that function. A uniform standard governing the permissible scope of collection should be avoided. In highly private settings, collection should be limited to the smallest amount of data required for a core function, and environmental information or biometric attributes unrelated to that function should be excluded. A somewhat broader scope of environmental data collection may be allowed in industrial and public settings, provided that the use of the data is strictly limited and does not extend beyond the function of the context to secondary use or algorithm training.

Biometric data also requires enhanced protection. Legislation should provide that biometric data may be processed only locally and may not be uploaded to the cloud unless the user’s separate written consent has been obtained. In highly sensitive settings such as homes and medical facilities, facial images, voiceprints, and other biometric data should be processed in real time on the device. Raw data should neither be uploaded to the cloud nor stored persistently. Where cloud processing is genuinely necessary, the user’s separate written consent should be obtained and end-to-end encryption applied throughout the process. The use of biometric data collected imperceptibly for algorithm training should also be expressly prohibited, reducing the risk of biometric information leakage at its source.

4.3. Full-Process Contextual Regulation

4.3.1. Ex Ante Market Access: Risk Prevention through Contextual Classification

A classification and tiering system grounded in scenario sensitivity and agent functionality should be implemented. Based on spatial sensitivity, application scenarios are stratified into three tiers: core private scenarios, semi-private scenarios, and public scenarios, each governed by differentiated sensor admission standards. Concurrently, based on the intended functions of the agent, products are categorized into four classes: companion and service robots, medical care robots, industrial collaboration robots, and public monitoring robots, each subject to distinct regulatory requirements.

A system of contextual privacy impact assessment should be established as a specialized form of the personal information protection impact assessment required by PIPL. Before a product enters the market, an enterprise should prepare a separate privacy risk assessment for each category of intended application rather than a general assessment of the product as a whole. The assessment should cover the sensitivity of the setting, the contextual necessity of data collection, risks throughout the data processing lifecycle, the effectiveness of technical safeguards, and emergency response mechanisms. The assessment report should be filed with the cyberspace authority. Products intended for highly sensitive settings such as medical facilities and homes should also undergo specialized review by independent experts, addressing privacy risks at their source.

Mandatory national standards should unify technical requirements for sensor precision, data retention, and local processing. In a home setting, for example, such standards could limit the resolution of a robot’s camera to 1080p, cap the data retention period at seven days, and require biometric data to be processed locally. These rules would convert the central requirements of contextual compliance into quantifiable and testable technical parameters. Separate mandatory limits could govern sensor precision, the retention period for raw data, and permissible methods of processing in settings with different levels of sensitivity.

4.3.2. Ongoing Regulation: A Multi-Actor Co-Governance Mechanism

Compliance responsibilities should be imposed throughout the chain of actors. Platform responsibility should reflect the special obligations imposed on important internet platforms by Article 58 of PIPL. Leading platform enterprises that provide operating systems or application distribution services for embodied AI should establish privacy compliance review mechanisms for products seeking access to their platforms, verify each product’s contextual compliance, and remove or require remediation of noncompliant products. Producers should also bear supply-chain management responsibilities. Suppliers of core sensors and algorithmic modules should be brought within a unified compliance management system, and producers should bear joint and several liability for privacy and security defects in upstream components, transmitting responsibility throughout the industrial chain.

A voluntary third-party certification system should also be established. Qualified professional bodies could certify products against contextual compliance standards and grant a uniform contextual compliance label to products that meet those standards, clearly indicating their approved applications and level of privacy protection. Certification results could be linked to government procurement, industrial support policies, and catalogs of green and smart products. Market signals would then guide consumers toward compliant products and create market incentives for enterprises to raise the level of privacy protection. This approach offers a representative means of coordinating flexible governance with binding regulation.

4.3.3. Ex Post Regulation: Liability Reform and Antitrust Enforcement

The rules governing liability for defective products under the Civil Code should be further specified. Where an embodied agent’s collection of data or interactive conduct seriously departs from the intended function and reasonable public expectations associated with the relevant setting, a design or algorithmic defect should be presumed and the producer should bear tort liability. A producer seeking exemption should prove that the conduct resulted from improper operation by the user or force majeure. Reversing the burden of proof in this way would address the evidentiary difficulties that algorithmic opacity creates for users, connect the ex post liability regime with the legislative framework for infringement liability, and provide courts with a clear and operational adjudicatory rule (Zheng 2022).

Regulation of market competition should respond cautiously to the risk of monopolization through contextual data. As embodied AI becomes widespread and offline contextual data accumulates at scale, the essential facilities doctrine should apply only when contextual data controlled by a dominant undertaking is irreplaceable and indispensable and a refusal to provide access would substantially exclude competition in the relevant market (Meng 2024). In those circumstances, access to anonymized and aggregated contextual data could be required on fair, reasonable, and non-discriminatory terms. Strict boundaries are needed to avoid excessive interference with commercial autonomy. Merger control should likewise treat the scale of offline contextual data as a central factor in assessing market competitiveness and prevent leading enterprises from acquiring a monopoly position through mergers that consolidate contextual data. This approach would maintain competitive market order while pursuing the dual objectives of data security and individual privacy protection.

4.4. Strengthening International Cooperation

The research and development of embodied AI, the division of labor across its industrial chain, and the contexts in which it is deployed are all highly globalized. Cross-border deployment of physical intelligent agents, cross-border data flows, and cross-border iteration of algorithms prevent any single national regime from forming a self-contained framework for privacy protection. Multilevel international cooperation is therefore necessary to build governance consensus, align rules and standards, and share governance capacity, thereby promoting a global balance between industrial innovation and privacy protection.

4.4.1. Participation in Global AI and Robotics Standard-Setting

China should participate deeply in the development of global standards for AI and robotics and promote the translation of contextual governance rules into international standards. Through multilateral standardization platforms such as the Artificial Intelligence Subcommittee of the National Information Technology Standardization Technical Committee (SAC/TC 28/SC 42) and the Ministry of Industry and Information Technology’s Technical Committee for the Standardization of Humanoid Robots and Embodied AI, China could propose incorporating several context-specific rules into future revisions of ISO/IEC 42001 on AI management systems. These rules include disabling continuous collection by default in home settings, providing prominent indications of collection status in public settings, and requiring local processing of core data in industrial settings. A uniform global compliance baseline would reduce the risk that divergent national standards become trade barriers and lower the cost of adapting embodied AI products to compliance requirements in overseas markets.

4.4.2. Consensus-Building through Multilateral Governance Frameworks

Multilateral governance frameworks can be used to build consensus and promote inclusive and balanced global rules. China can rely on its Global AI Governance Initiative to advocate an approach based on risk classification, contextual adaptation, and the coordination of development with security in such multilateral mechanisms as the United Nations AI Advisory Body, the United Nations Secretary-General’s High-level Panel on Digital Cooperation, and the G20 Digital Economy Ministers’ Meeting. Cooperation should focus on common risks, including intrusion into privacy in physical space and continuous imperceptible collection by embodied AI, and should seek shared governance rules.

4.4.3. Bilateral Compliance Cooperation and Capacity Building for Developing Countries

Bilateral industrial compliance cooperation and capacity building for developing countries should be deepened through a differentiated framework. With economies that have mature digital governance rules, including the European Union, China could pilot bilateral mutual recognition of privacy compliance certification for embodied AI. Third-party privacy testing results for smart-home products and service robots could be coordinated with the European Union’s CE certification, reducing duplicative compliance costs for enterprises. Under the framework of the Global Development Initiative, China could also support capacity building for embodied AI governance in developing countries. Technical assistance, professional training, and the sharing of regulatory sandbox experience would help those countries develop privacy rules suited to their own stage of development and narrow the global divide in digital governance capacity.

5. Conclusions

The development of embodied AI is moving artificial intelligence from informational space into the physical world. It is reshaping production and everyday life while generating privacy risks of an unprecedented kind. The central task of privacy governance for embodied AI is a shift in regulatory logic from static data control to dynamic contextual regulation. The theory of contextual integrity requires privacy to be assessed within specific contexts of physical interaction, with particular attention to whether information processing conforms to reasonable expectations in the relevant setting and to its actual effects on the rights and interests of natural persons. China has already formed an initial governance architecture composed of foundational statutes, departmental rules, technical standards, and local pilot programs.

Future legislation should move beyond static data control through a layered legislative strategy that establishes contextual compliance as a legal principle, clarifies the instrumental status of intelligent agents and the boundaries of responsibility among relevant actors, and subjects biometric data to a strict principle of local processing. Regulatory implementation should create a full-process mechanism encompassing tiered ex ante access, multi-actor ongoing governance, and innovative ex post liability. Privacy obligations should be embedded in product design, market circulation, and every stage of contextual application. A dynamically updated system of technical standards should translate abstract legal principles into quantifiable and enforceable engineering norms. By balancing incentives for industrial innovation with individual privacy protection, China can provide stable legal expectations for the sound development of embodied AI and may also contribute a constructive institutional model to the formation of global AI governance rules.