世卫组织-人工智能健康伦理与治理：大型多模式模型指南（英）-2024-98页_1mb

Summary

The World Health Organization (WHO) guidance addresses the use of large multi-modal models (LMMs) in health care, drawing on existing ethical principles for AI and proposing recommendations for governance and risk mitigation. LMMs, a form of generative AI that can process multiple data inputs and produce diverse outputs, are already being used in health diagnostics, patient care, medical education, and drug development. However, they introduce novel risks such as inaccurate responses, bias, hallucination, privacy concerns for individuals and data workers, cybersecurity vulnerabilities, and potential societal harm.

The guidance is structured around the AI value chain, examining risks and proposed measures at three stages: design and development, provision, and deployment. During development, developers must implement measures to address bias, privacy, data quality, and ethical transparency; provision involves government enforcement of regulations to govern the use of LMMs in health care; and deployment requires ongoing monitoring and adherence to safety protocols, with accountability mechanisms to mitigate risks.

Key recommendations include:

• Governance measures require governments to play a central role in formulating and enforcing laws, ethical principles, and international standards for LMMs.
• Mitigation strategies focus on requiring transparency (e.g., labeling AI-generated content), protecting data privacy, and promoting human oversight.
• Liability: introducing "presumption of causality" rules ensures accountability for individuals harmed by LMM use, while preventing overly narrow liability frameworks.
• International governance is emphasized to ensure equitable global participation and prevent a "race to the bottom" in regulations.

Several risks specific to health care are highlighted, including potential automation bias, skills degradation, and concerns about the quality of human-AI interactions, reinforcing the need for clear guidelines and ongoing impacts assessments. LMMs also have environmental impacts, such as high energy consumption and significant water footprints, suggesting the need for energy-efficient design and global environmental standards.

Overall, the guidance supports a human rights-based approach to the governance of LMMs, stressin that ethical principles should guide their use to maximize health benefits while safeguarding human rights and minimizing risks.