We have several working groups to proceed with the review of design guidelines and the preparation of sample data, categorized by usage scenarios, content, and countermeasures, with the aim of establishing a full-scale benchmark.
Artificial intelligence (AI) has the capacity to generate responses that reflect the social biases present in its training data. As AI systems are increasingly deployed across diverse application domains, there is a growing need for objective methods to evaluate whether they can produce socially appropriate and non-discriminatory responses in a wide range of social contexts. This working group is focusing on the construction of a social bias evaluation dataset grounded in statutory laws.
This working group addresses the risk that interactions with AI themselves may have a negative impact on users and those around them. The impact of such interactions inherently varies depending on the user’s attributes, vulnerabilities, and other factors. With children as the target users, this working group is working to develop benchmarks to ensure that “AI can be used safely by children.
This working group is to develop a benchmark to address input and output errors in LLMs, such as fake news and hallucinations. We will pursue two lines of work: collecting real-world examples of false and misleading information from social media, and generating synthetic misinformation QA data to assess whether LLMs can provide accurate, factually grounded responses. Through these efforts, we aim to advance the evaluation of LLM robustness against mis- and dis-information.
When utilizing AI, challenges arise that depend on the specific application domain. This working group is to define safety evaluation criteria across various fields, identify issues specific to each domain as well as common challenges, and ensure that the benchmarks developed will be useful across a wider range of fields.
Healthcare is a field that demands particularly high safety standards, and the effective use of AI requires careful consideration of the unique context of Japan’s healthcare system, as well as its customs and guidelines related to health. This subgroup is working to develop benchmarks for the healthcare sector that reflect these conditions specific to Japan.
The security of the Japanese language must be assessed using the Japanese language itself. We will develop a benchmark to evaluate the robustness of models against jailbreak attacks, taking into account aspects unique to the Japanese language—such as inconsistencies in notation, honorifics, and role-playing in practical business contexts—that are not covered by overseas benchmarks alone.
Among AI-related data leaks, we are focusing on data leaks occurring in AI-generated responses with search tools (RAG) , and we are currently developing a benchmark to address this issue. Data leaks caused by in-house RAG systems have already become a problem, and we believe this issue will become even more prevalent as the delegation of authority to AI continues to expand. Given these circumstances, we aim to develop a benchmark with a view toward practical application.
With the goal of ensuring the safe and secure use of AI in the field of robotics, the Robotics SWG—established as a thematic subcommittee of the AISI Business Demonstration Working Group—will conduct demonstration evaluations using actual robots and develop a guide on AI safety evaluation criteria and related evaluation methods.
The Evaluation Platform Subcommittee is working to design and build a platform that utilizes benchmark data created by each subcommittee to evaluate the safety of LLMs from multiple perspectives. With the goal of creating an environment where LLM users and developers can easily and continuously evaluate safety, the subcommittee is striving to balance evaluation quality with usability.
The objective of this subcommittee is to develop benchmark datasets, evaluation environments, and evaluation protocols for the objective assessment of AI models’ cybersecurity capabilities and associated risks. While AI is utilized for vulnerability analysis, threat detection, log analysis, incident response, and vulnerability remediation support, it also possesses capabilities that could potentially be repurposed for malicious purposes; therefore, a foundation is needed to measure these capabilities and risks in a reproducible and secure manner. This subcommittee will evaluate vulnerability detection capabilities, the potential for inferring exploitation procedures and generating malicious code, and the potential for the misuse of malware-related knowledge under conditions where risks are safely neutralized and abstracted, with the goal of establishing evaluation metrics that contribute to the safe use of AI and risk mitigation.
The purpose of this subcommittee is to develop benchmark datasets, evaluation environments, and evaluation protocols for assessing the safety and security of agent-based AI models that involve tool usage, environmental manipulation, multi-stage planning, and iterative decision-making. Since agent models are capable of performing actual operations on external tools and systems, evaluation must encompass not only their outputs but also their behavioral processes, access control, and impact on the environment. This subcommittee will conduct reproducible evaluations of task execution capabilities under conditions close to real-world environments, the appropriateness of tool usage, resistance to unintended behavior, deviations, and instruction injection, the suppression of operations exceeding authorized limits, and the prevention of inappropriate access to or transmission of confidential information. Through these efforts, we will work to establish evaluation metrics for agent-based AI that balance utility with safety and security.
To Be Displayed