Responsible Artificial Intelligence Policy
Release time:
2026-07-24
1. Purpose and Principles
Livzon Pharmaceutical Group Inc. (hereinafter referred to as the “Company” or “Livzon Pharmaceutical”) fully recognizes the significant potential of artificial intelligence (AI) technologies across the entire lifecycle of pharmaceutical activities, including drug discovery, clinical research and development, process optimization, manufacturing, quality control, supply chain management, and patient support. At the same time, the Company is aware of associated risks and challenges related to data security, algorithmic bias, decision-making transparency, patient safety, and environmental impact.
To establish a trustworthy, responsible, and auditable AI governance system, and to ensure that the development, deployment, and use of AI technologies adhere to the principles of legality and compliance, safety and controllability, ethical appropriateness, transparency and explainability, fairness and non-discrimination, and environmental sustainability, the Company formulates this Policy. This Policy aims to effectively protect patient privacy, data security, and intellectual property rights, and to reduce AI-related energy consumption and carbon emissions. It is established in accordance with the Personal Information Protection Law of the People’s Republic of China, the Data Security Law of the People’s Republic of China, the Cybersecurity Law of the People’s Republic of China, the Interim Measures for the Administration of Generative Artificial Intelligence Services, relevant regulations of the National Medical Products Administration (NMPA), the WHO Guidance on Ethics and Governance of Artificial Intelligence for Health, and the core principles of the European Union Artificial Intelligence Act (EU AI Act).
2. Scope of Application
This Policy applies to all departments, employees, and contractors of the Company and its subsidiaries (collectively referred to as the “Group”), as well as all activities conducted in the name of the Company. The AI systems covered include all AI models, algorithms, tools, applications, and generative AI systems developed in-house, procured externally, co-developed, or used by the Company (whether deployed on-premises, in private clouds, or on third-party platforms), as well as AI functional modules embedded in existing business systems.
3. Governance Framework and Approval
The Group has established a three-tier AI governance structure: the Board of Directors serves as the highest approval and decision-making body for this Policy and assumes ultimate responsibility for AI governance, reviewing AI risk and compliance reports at least once a year; the ESG Committee performs day-to-day supervisory duties; and the AI Governance Working Group is responsible for policy implementation, risk assessment, training supervision, and continuous improvement. Each department and subsidiary shall designate an AI responsible person to implement daily compliance management. AI governance responsibilities are incorporated into the ESG-related performance evaluations of relevant departments and key positions.
4. Core Policies
4.1 Data Privacy Protection
Throughout the entire process of AI development and use, the Company strictly complies with the Personal Information Protection Law of the People’s Republic of China, the Data Security Law of the People’s Republic of China, the Cybersecurity Law of the People’s Republic of China, and other applicable laws and regulations, as well as data protection regulations in all operating regions, including the EU General Data Protection Regulation (GDPR). The principles of legality, legitimacy, necessity, and data minimization, as well as purpose limitation, shall be followed. Only the minimum dataset directly relevant to the AI application scenario shall be collected. Explicit authorization must be obtained prior to collection. Use beyond the authorized scope, or disclosure, sale, or sharing of data with unauthorized third parties, is prohibited. In automated AI decision-making scenarios, the data subject’s rights to be informed, to access, to rectify, to erase, and to object shall be safeguarded in accordance with the law. Personal data used for AI training shall preferentially undergo de-identification or anonymization. For sensitive data such as patient health information, a Data Protection Impact Assessment (DPIA) must be completed and approved by the relevant departments. When transferring personal data to third-party AI service providers, a Data Protection Agreement (DPA) must be signed, clearly specifying security measures, audit rights, and liability for breach.
4.2 System Cybersecurity
A cybersecurity protection system covering the full AI lifecycle shall be established to guard against threats such as malicious attacks, data tampering, model theft, adversarial sample attacks, data poisoning, and prompt injection. Multi-layered protections (including firewalls, WAF, IDS/IPS, and API security gateways) shall be deployed. Transmission and data-at-rest encryption, least-privilege access control, and multi-factor authentication shall be implemented. Complete audit logs of critical operations shall be retained. Regular penetration testing, vulnerability scanning, and specialized AI security assessments shall be conducted, with critical systems assessed at least once every six months. An AI cybersecurity and data breach emergency response plan shall be formulated and exercised at least once a year. Security and compliance due diligence shall be performed on third-party AI suppliers. Preference shall be given to suppliers holding certifications such as Classified Protection of Cybersecurity or ISO 27001. Contracts must include security clauses, audit rights, and exit mechanisms.
4.3 Human Oversight and Human Intervention
The principle of “human agency and oversight” shall be upheld. In scenarios involving personal safety, patient health, drug quality, significant regulatory decisions, or high-risk medical applications, qualified professionals shall retain final decision-making authority, intervention rights, and review rights. Critical AI systems must incorporate predefined Human Override Points. When AI outputs are anomalous, highly uncertain, or involve patient safety, the system shall be promptly paused, overridden, or subject to mandatory human review. Fully autonomous decision-making AI systems that cannot be effectively intervened by humans are prohibited from deployment in high-risk healthcare domains.
4.4 Transparency, Explainability, and AI Identification
Tiered transparency obligations shall be fulfilled for different stakeholders. For patients and the public, prior notice shall be provided that they are interacting with AI, and the decision basis, confidence level, and limitations shall be explained in plain language. For healthcare professionals, technical explanations shall be provided, including algorithm principles, key input features, training data distribution, model performance metrics, and known limitations. For regulatory authorities, complete technical documentation shall be provided (including system architecture, training data summary, validation reports, and bias detection records).
An AI-generated content and decision identification system shall be established. In scenarios involving critical information or key business activities, AI-generated text, images, and other content shall be clearly labeled as AI-generated. For non-critical scenarios or general auxiliary content, business units may implement identification flexibly according to actual risk levels. Complete audit logs shall be retained for internal AI applications to support regulatory inspections and post-event traceability.
4.5 Anti-Bias and Fairness
Algorithmic bias shall be systematically identified and mitigated at all stages of AI model development, training, validation, and deployment to ensure that no discriminatory impact occurs across dimensions such as gender, age, ethnicity, geography, or disease type. Training data must be diverse, representative, and balanced. AI applications targeting the Chinese market must adequately cover major subgroups of the Chinese population. A bias audit mechanism combining quantitative methods (subgroup testing with statistical fairness metrics) and qualitative methods (multidisciplinary expert review) shall be established and executed prior to model launch, upon major updates, and during regular operations, to ensure that AI decisions meet the core requirements of fairness and impartiality.
4.6 Accountability Mechanism
The principle of “whoever develops or procures is responsible, whoever deploys or uses is responsible, and whoever reviews is responsible” shall be followed, with clear definition of the responsibilities of parties involved at each stage of the AI lifecycle. AI developers shall ensure model compliance and record the full development process. AI deployers shall be responsible for operational monitoring and risk handling. AI users shall bear ultimate professional responsibility for business decisions made based on AI outputs. An accountability mechanism for AI-related incidents shall be established: entities that intentionally violate rules or commit serious negligence resulting in patient harm, data breaches, or regulatory penalties shall be held accountable in accordance with laws and regulations. For damages caused by defects in AI systems, corresponding compensation and remedy mechanisms shall be established to protect the legitimate rights and interests of affected parties.
4.7 Application Boundaries and Prohibited Practices
All AI systems shall undergo risk level assessment prior to deployment, with control measures determined according to the risk level. Product/service descriptions and user agreements for AI systems must clearly state functional boundaries, intended uses, known limitations (such as hallucinations, data drift, and uncertainty), contraindicated scenarios, and user obligations. Exaggeration or false advertising of AI capabilities is prohibited. These boundaries shall be translated into executable control rules, including access permission for AI use scenarios and privileges, risk interception mechanisms during AI operation, privilege restrictions on AI function calls, and escalation procedures for boundary anomalies. Annual verification shall be conducted to confirm that AI systems do not operate beyond approved boundaries.
Strict compliance with regulatory red lines shall be maintained. The development, deployment, or use of the following AI systems is explicitly prohibited: (1) AI systems that deploy subliminal, manipulative, or deceptive techniques that substantially distort individual behavior in a manner that is likely to cause harm; (2) AI systems that exploit the vulnerabilities of specific groups (such as age, disability, or socio-economic situation) to cause harm; (3) AI systems that perform social scoring based on social behavior or personal characteristics resulting in unfair treatment; (4) Real-time remote biometric identification systems used in publicly accessible spaces for law enforcement purposes (except where explicitly authorized by law); (5) Biometric categorization systems that infer sensitive attributes such as race, political opinions, or religious beliefs, as well as emotion recognition systems improperly used in workplaces and educational institutions; (6) AI systems that generate false medical information, clinically unverified prescriptions, or regulatory documents without effective human review; (7) Other AI applications prohibited under Chinese law and international norms to which the Company is committed. Items (1) to (5) above are based on the prohibited practices set out in Article 5 of the EU Artificial Intelligence Act (EU AI Act).
4.8 Sustainable Development and Green AI
AI computing resources shall be planned in an integrated manner, with priority given to low-energy and high-efficiency methods of deployment and use. Lightweight model architectures shall be preferentially adopted to reduce energy consumption and carbon emissions during training and inference stages. Elastic scheduling and resource sharing shall be implemented on owned or hosted AI computing platforms to improve utilization rates. Preference shall be given to suppliers with green data center certifications and lower Power Usage Effectiveness (PUE). AI energy consumption and carbon footprint shall be included within the scope of the Company’s ESG report disclosures, and the gradual establishment of AI carbon emission accounting methods is encouraged. The positive application of AI in green process route design, energy management optimization, and supply chain carbon reduction is encouraged. Relevant projects’ contributions to the Group’s sustainable development goals shall be measured, and an assessment indicator system for the sustainable development contributions of AI projects shall be established and quantified for regular disclosure, providing data support for decision-making and demonstrating the fulfillment of responsibilities to stakeholders in a timely manner.
5. Control of High-Risk AI Capabilities
Strict controls shall be applied to high-risk AI capabilities that may have a significant impact on individual health, safety, or fundamental rights. High-risk AI systems include, but are not limited to: AI systems used for clinical decision support, treatment recommendations, and prognosis assessment; AI systems used for critical decision-making in pharmaceutical quality control; AI systems used for the detection of adverse drug reactions or safety signals; AI systems used for clinical trial inclusion/exclusion determination and endpoint assessment; and AI systems involving biometric processing such as facial recognition or emotion analysis (except for legitimate purposes such as security protection). Prior to deployment, such systems must complete risk assessment and obtain review by the AI Governance Working Group. A high-risk AI capability inventory and authorization control matrix shall be established, allowing only authorized personnel to invoke relevant capabilities. The deploying entity, usage scenario restrictions, and operational boundaries of such technologies shall be clearly defined. Regular monitoring of operational effectiveness and compliance reviews shall be conducted.
6. AI Model Performance Maintenance and Monitoring
A full-lifecycle performance monitoring and continuous maintenance mechanism for AI models shall be established to ensure that AI systems continuously meet performance, safety, and compliance requirements. Continuous or periodic monitoring of AI models in the production environment shall be implemented, covering core dimensions including model performance degradation, data/concept drift, bias drift, anomalous output rates, and abnormal resource consumption. When monitoring indicators trigger warning thresholds, the operations responsible person shall initiate diagnosis within the prescribed time limit and take targeted corrective measures, such as retraining the model with new datasets or adjusting model operating parameters.
7. AI Decision Appeal Mechanism
A smooth and fair AI decision appeal mechanism shall be established to ensure that individuals and organizations affected by AI decisions have access to channels for raising objections and seeking remedies. This mechanism applies to situations including: significant disputes regarding AI-assisted clinical diagnosis or treatment recommendations; AI decisions involving the processing of personal health information that infringe upon legitimate rights and interests; automated AI personnel decisions that produce material adverse effects; and AI outputs that lead to disputes over pharmaceutical quality or adverse event determinations. Appellants may submit appeals through the Company’s official website, the designated hotlines (0756-8135299 or 0756-8135948), or written correspondence. Investigation and assessment shall be completed within 30 calendar days, and the handling conclusion shall be provided in writing within 10 working days. Where an AI decision is verified to be erroneous, remedial measures such as correcting the decision, restoring rights and interests, improving the system, or adjusting processes shall be taken. Appeal data shall be summarized and analyzed annually to identify systemic issues and drive continuous improvement.
8. Training and Capability Building
The Group integrates the AI training system into its overall development strategy and continuously improves the AI governance mechanism. Differentiated and hierarchical training programs shall be implemented according to job responsibilities and risk levels. Training content shall cover compliance requirements, system security, ethical principles, and responsible AI application. The AI Governance Working Group shall organize annual assessments of training effectiveness and continuously optimize training content based on the assessment results. For problems and deficiencies identified during AI use, the Company commits to taking timely corrective measures and promoting the continuous iteration and improvement of the governance mechanism.
In addition, the Group actively expands diversified learning channels, including but not limited to establishing AI communities and organizing thematic seminars and practical exchange activities, to enhance the AI literacy and compliant application capabilities of all employees from multiple dimensions, and to promote the practical implementation and value realization of AI capabilities across business scenarios.
9. Policy Implementation, Review, and Revision
This Policy shall be reviewed and approved by the Board of Directors or its authorized ESG Committee before release, and shall take effect from the date of release. The ESG Committee assumes supervisory responsibilities. A comprehensive review shall be conducted at least once a year, or the Policy shall be revised in a timely manner upon significant changes in laws and regulations, evolution of AI technology, business adjustments, updates to disclosure requirements, or occurrence of major AI security incidents. The Company encourages the exploration of engaging third-party institutions to assess and certify the AI management and governance system, and to conduct verification against relevant domestic and international responsible AI practice standards such as ISO 42001, to validate the compliance and effectiveness of the system. Relevant verification results may serve as reference for the revision of this Policy and the optimization of the AI governance system.
The final interpretation and revision rights of this Policy rest with the Board of Directors.
Previous Page