AISI Formulates 'AI Safety Evaluation Perspectives Guide in the Healthcare Domain'

The Healthcare Sub-Working Group (hereafter, SWG), one of the Business Verification Working Groups (hereafter, Business Verification WG) of the AI Safety Institute (AISI, Director: Akiko Murakami), has formulated the "AI Safety Evaluation Perspectives Guide in the Healthcare Domain" (hereafter, this Guide) to accelerate the safe social implementation of generative AI technology in the healthcare domain.

This Guide presents specific evaluation methods for practitioners, reflecting the sensitivities and risks unique to medical care and healthcare, with reference to the AI safety evaluation perspectives guide formulated by AISI.

* IPA handles a part of the secretariat functions of the AI Safety Institute.

Background and Purpose In recent years, generative AI has brought significant innovation to the healthcare field, such as streamlining doctors' operations and supporting patient communication. On the other hand, there are numerous challenges specific to the healthcare domain, including the risk of health hazards due to hallucinations (generation of misinformation), advanced privacy protection, and ensuring security.

In the Basic Plan for Artificial Intelligence approved by the Cabinet in December 2025, the promotion of AI development, verification, introduction, and social implementation in the medical and healthcare fields was explicitly stated as a specific initiative. Furthermore, as discussed at the "Hiroshima Global Forum for Trustworthy AI" held in January 2026, the realization of "Trustworthy AI" has become an important issue in the international community as well.

Recently, the Healthcare SWG has been operating with the aim of developing an environment where businesses can ensure safety from the development and design stages, balancing business value with safety and security, and has formulated this Guide.

Main Features of the Guide This Guide has the following features so that even companies with few experts can easily utilize it.

### 1. Evaluation Points for Each of the 5 Phases Along the AI Lifecycle It summarizes the points of what should be evaluated in each of the five phases along the AI lifecycle: product design, model selection, product implementation, product verification, and product introduction/operation.

### 2. 10 Multifaceted Evaluation Perspectives and Specific Risk Assumptions Based on the AI safety evaluation perspectives guide, it specifically lists 10 items regarding what perspectives are necessary when applying them to the healthcare domain. In addition, it also describes the risks expected by neglecting evaluation from those perspectives, such as "providing incorrect or dangerous information regarding medical care and health could bring direct harm to the life and health of patients." These cover everything from universal risks to those unique to the healthcare domain, and are also useful for risk management when utilizing AI.

Furthermore, a practical guide containing specific examples (prompts, agent skills, etc.) to enable the implementation of AI safety evaluations based on the evaluation perspectives of this Guide is also scheduled to be released in the future.

Participating Organizations in the SWG - Ubie, Inc. (SWG Leader), Awarefy Inc., Ajinomoto Co., Inc., SB Intuitions Corp. - CMIC HOLDINGS Co., Ltd., SherLOCK, Inc. - Japan Digital Health Alliance (JaDHA), MICIN, Inc., The Tokyo Foundation for Policy Research - Mitsubishi Research Institute, Inc. (SWG Secretariat)

Future Outlook The Healthcare SWG positions this Guide as a practical handbook for realizing trustworthy AI, and will update it as appropriate in response to rapidly changing generative AI technologies, social conditions, and international regulatory trends. Through this, we will contribute to the safe social implementation of AI in the healthcare field and the creation of sustainable business value.

Reference Information AISI Healthcare SWG Introduction Video (IPA YouTube Channel) * IPA handles a part of the secretariat functions of the AI Safety Institute Secretariat.

Inquiries Regarding this Matter AI Safety Institute Secretariat: Futamura, Kato, Maeda Email: [email protected]