This page has been translated using TexTra by NICT. Please note that the translation may not be completely accurate.If you find any mistranslations, we appreciate your feedback on the "Request form for improving the automatic translation ".

The 6th Advanced AI Utilization Advisory Board

Last Updated:

Outline

For an overview of the Committee, please see Advanced AI Utilization Advisory Board .

Event Information

Date:
Tuesday, September 29, 2026 from 11:00 a.m. to 1:00 p.m.
Location:
Office of the Prime Minister's Office - Online

Proceedings

  • 1. Opening
  • 2. Proceedings
    • 1. Outline of Periodic Report on generative AI System by Each Ministry and Agency
    • 2. Report on the Second Meeting of the Working Group
    • 3. Report of the Third Meeting of the Working Group
    • 4. Trends in generative AI in Japan and other countries
    • 5. Explanation from the Chai Shan
    • 6. Proposed Considerations for the Enhancement of the Guidelines for the Acquisition and Utilization of generative AI for the Evolution and Reform of Public Administration
  • 3) Closing

Data

Reference Documents

Summary of proceedings

At the beginning, Mr. State Minister for Digital Transformation Shiozaki stated that at a time when society and administration are facing a major turning point with the rapid development of AI, this plenary session is important as a guiding principle for connecting the power of AI to the evolution of administration and the improvement of services for the people.

(1) Outline of Periodic Reports on generative AI System by Each Ministry and Agency

The secretariat submitted Appendix 1 as a report on the results of an analysis of the state of generative AI utilization by the various ministries and agencies.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • Yoshinaga: Member of the Committee: In relation to "(ii) Risk of leakage of Information designated" confidential "Information that requires confidentiality" in "Proposed Matters for Consideration II Reinforcement of Unique Response Measures Accompanying Expansion of AI Agent Targets" on page 14 of the As described on page 7 of Exhibit 1, AI is increasingly being used to create scenarios for political simulations. As the autonomy and automation of AI increase, risks also increase. Therefore, it is necessary to pay close attention to future trends.
    • Secretariat: As for the examples of policy scenario study using generative AI, it is conducted based on the approach of "Human in the Loop" as a general rule. In the other AI agent use cases currently under investigation, it is also assumed that humans make decisions and confirm. It is necessary to continue to proceed while considering the balance between use and risk.
  • Yoshinaga: Member of the Committee: In relation to "(ii) Risk of leakage of Information designated" confidential "Information that requires confidentiality" in "Proposed Matters for Consideration II Reinforcement of Unique Response Measures Accompanying Expansion of AI Agent Targets" on page 14 of the Material 1, is it correct to understand that they are used for program testing, etc.? We would like to ask for specific applications.
    • Secretariat: The implementation of autonomous AI agents is introduced, for example, on page 9 of Appendix 1. However, there are cases where multiple tasks such as information collection and analysis are executed in cooperation, not limited to program testing.
  • Nabatame member: There is a viewpoint that it is difficult to make a judgment on the governance of We appreciate that Document 1 has been updated on a regular basis. As a trend, there has been a small but increasing number of cases where chatbots are used by a single government agency or across multiple government agencies. We think this trend is natural. As for chatbots mentioned in Reference 1, page 6, if customers' inquiries about products and services are collected in an integrated manner, each ministry and agency will need a system that can respond to a wide range of inquiries. For this reason, it is necessary to build an integrated and comprehensive governance system, as well as to monitor it closely. In addition, as customers increasingly use AI agents to select products and services, each business operator has started initiatives for "AI optimization," which optimizes customer experience based on the selections made by AI agents. In light of these environmental changes, it is necessary to pay close attention to how chatbots should be utilized. In Singapore, the beta version of "chat. gov. sg," a government-official AI chatbot service for administrative information guidance, has been deployed, and inquiries to all ministries and agencies have been centralized. In the future, when the Government of Japan considers similar services, it is necessary to discuss how to ensure unity among the ministries and agencies as well as the Government as a whole.

(2) Advanced AI Utilization Advisory Board 2nd Working Group Implementation Report

The Secretariat and members reported on the outline of the Second Meeting of the Working Group in Appendix 2, and members asked questions and made comments.

(3) Advanced AI Utilization Advisory Board, Working Group, Third Meeting Implementation Report

The secretariat and members reported on the outline of the 3rd Working Group meeting in Appendix 3.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • Yoshinaga: Member of the Committee: In relation to "(ii) Risk of leakage of Information designated" confidential "Information that requires confidentiality" in "Proposed Matters for Consideration II Reinforcement of Unique Response Measures Accompanying Expansion of AI Agent Targets" on page 14 of the In "Discussion Point (2) Expansion of utilization and development in business" on page 6 of Document 3, there is a description that "It is said that personal profiling is prohibited, but the work related to clearance is also increasing in the government." In the United States, the Federal Trade Commission (FTC) published a report on big data in the early 2000s. It was pointed out that profiling may continue after disadvantageous or discriminatory treatment occurs. This point needs to be carefully considered.
    • Secretariat: profiling, it is possible to investigate individuals, such as employees of information management departments, but it is controversial whether profiling should be conducted in GENAI or elsewhere. On-premises systems and systems built by each ministry and agency are generally used, and it is necessary to use them depending on the nature of the business.
  • Nabatame member: There is a viewpoint that it is difficult to make a judgment on the governance of process, we believe that appropriate governance has been taken at this stage.
  • Torisawa: : I think the boundaries between governance, technology, and AI will become blurred. It is desirable to have people who are familiar with AI agent technology participate in this discussion. It is necessary to proceed with the consideration based on technological trends, including the use of state-of-the-art tools as a procurement requirement.

(4) Trends in generative AI in Japan and other countries

The Secretariat reported the trend of generative AI in Japan and other countries in Appendix 4.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • Nabatame member: There is a viewpoint that it is difficult to make a judgment on the governance of Material 4 Regarding the discussion of operational efficiency improvement and value creation in "AI Usage by AI Introduction Effect", if only efficiency improvement is pursued, the organization may move in a biased direction. It is necessary to make it a common understanding that value creation is equally important.
    • Secretariat: While it is important to accumulate improvements in efficiency at the individual task level, we believe it is also important to improve the productivity of the entire organization. Even if individual processes become more efficient through the use of generative AI, it should not be the case that the entire organization becomes less efficient. In future initiatives, it will be necessary to effectively utilize the resources freed up by the use of generative AI and to reassign employees to higher value-added duties. We will continue to consider this as a challenge.
    • Nabatame member: There is a viewpoint that it is difficult to make a judgment on the governance of Finally, how to improve the organizational capacity of each ministry and agency becomes an important issue. The company is also working based on the idea that "investment in AI is investment in people." I understand that the government is also working based on the same idea.

State Minister for Digital Transformation Shiozaki gave the following remarks before leaving the meeting: It is meaningful to consider the use of AI in various administrative activities, including AI use in child guidance centers. In light of the rapid pace of technological progress, it is also necessary to consider the ideal cycle of AI use and development in governments.

(5) Explanation from a Chai Shan

The Chai Shan reported on the JDLA AI Governance Certification System-C certification in Appendix 5.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • A member of KITAMURA: ISO/IEC42001 and C-Certification are not mutually exclusive. It is important to detect changes in data models, AI, usage environments, etc., and how fast we can make decisions and resynchronize updates. I call this "Resync." As Chai Shan members pointed out, it is even more difficult when it crosses ecosystems and supply chains and also crosses countries. Kitamura himself has proposed a lightweight framework called "Minimum Viable Collaboration Mechanism (MVCoM)." However, the issue is how to utilize C-Certification from the viewpoint of not only certification acquisition but also resynchronization.
    • Chai Shan Member: C certification should not be regarded as the final stage, but should be linked to further efforts such as ISO/IEC 42001. The synchronization of these systems is also an issue for the future, and we will continue to consider the consistency between systems and the response of the entire supply chain.
  • Sakuma: systems, I understand that C certification is positioned as a baseline when embarking on AI governance efforts. Including the "AI Governance Navi" introduced in the 5th Advanced AI Utilization Advisory Board, I believe that the structure will develop into organizational governance certification such as ISO/IEC42001.
  • Naganuma: As mentioned in the explanation, private companies in particular tend to put emphasis on the development of rules, but in AI governance, it is important to incorporate it into operation. I think that the fourth point of the certification standards of C certification, "Risk response," described on page 4 of Appendix 5, will be the core. We believe that the concept of quickly launching in a relatively short period of two to three months will continue to spread. On the other hand, in order to cover the entire supply chain and spread it to various organizations, it takes a certain amount of effort to cover all of 1 to 4 of C certification. I would like to know what kind of confirmation process you assume, such as whether you will put it through the certification process in some way or apply by yourself, when updating the items stipulated in "Risk Response" after certification.
    • Chai Shan Member: As a premise, the content of certification criteria 3 "risk assessment" and 4 "risk response" differs from company to company. Even if the same AI is used, we believe that the concept and response to risks will differ depending on each company's attitude toward risks. Therefore, rather than examining individual risks in detail, we expect that the examination will be conducted from the perspective of whether the function of risk assessments and risk responses is implemented appropriately, and whether it is not unreasonable. In addition, it is assumed that each company will periodically conduct risk assessments and review risk responses. In light of technological development, the system is scheduled to be renewed every two years, but we will continue to consider the specific method of renewal examination.
    • Naganuma: will continue to be considered.

(6) Proposed Considerations for Enhancing the Guidelines Concerning the Procurement and Utilization of generative AI for the Evolution and Reform of Public Administration

The Secretariat reported a draft agenda for improving the Guidelines in Appendix 6.

Questions and Opinions

The main questions and opinions of the attendees are as follows.

  • A member of KITAMURA: as a whole, I think it will be necessary to refactor existing governance. Currently, it is assumed that humans read regulations and make judgments, but considering that AI agents will make judgments and actions in the future, we need to shift from governance where humans read regulations and respond to them to governance where AI can interpret and execute them. I believe that Japanese experience and track record in the manufacturing ecosystem and multi-layered supply chains will be an advantage that will allow us to have a global presence. I believe that it is necessary to consider governance from a higher layer. What do you think?
    • Secretariat: : The governance approach is as you pointed out. We believe that a risk-based governance approach will be necessary, taking into account the autonomy of AI agents, the irreversibility of tasks, and the impact of execution. We will continue to study this.
    • Secretariat: people should play and the roles that AI agents should play. It is also expected that how government officials will build its governance will change significantly. Rather than leaving everything to AI agents, we will consider ideas such as "Human in the Loop" and "Human in the Lead," while taking into account discussions at the Liberal Democratic Party's Digital Society Promotion Headquarters and other bodies.
    • A member of KITAMURA: If additional actions are taken on existing governance, there is a risk of accumulating governance liabilities. We therefore request the ARRC to continue such consideration.
  • A member of KITAMURA: An agent may take unintended actions, such as exceeding its granted authority or overriding its goals. You have explained the need for a kill switch. However, without determining how far to stop and who should stop, it may be difficult to make decisions during actual operation. When starting such operation, it is necessary to develop some kind of control measure. Please tell us what considerations can be shared at this point.
    • Secretariat: : The kill switch is an issue related to the control of AI agents. We believe that the standard will be a mechanism that centrally manages gateways such as MCP, which is becoming a de facto operation in the industry, and makes it possible to block them as necessary. As for the judgment criteria in terms of operation, one possible operation method is to monitor the dynamic operation status in addition to human confirmation, and when signs of risk are confirmed, suspend it and see how it is. We consider that recovery procedures after suspension are a necessary function when operating a highly available system. On the other hand, looking at the utilization status of generative AI at each ministry and agency, there are only a limited number of examples of utilization of AI with a high degree of autonomy at present, and we would like to consider the introduction based on the need.
    • Chairman: Kadobayashi, AI With regard to the autonomy of agents, a high-horsepower frontier model that makes inferences on its own has emerged. Discussing such a model in the same framework as traditional agents that move in accordance with human prompts is a matter of concern internationally. The possibility of grasping the thinking process through the "chain of thought" is also an important issue. In the "looped transformer" type models such as Claude Mythos, there is a risk that thinking cannot follow. The balance between improving accuracy and reducing observability is an international debate. When using AI agents for the preparation of administrative documents, etc., it is also necessary to discuss the degree of autonomy or automation separately. The discussion of the kill switch is a point that is also related to the discussion of the frontier AI.
    • A member of KITAMURA: : As for the definition of AI agent and agent AI, the OECD has indicated one approach, and AI Guideline for Business has also defined it, but it is a very broad concept, and there is a risk that it will be interpreted in a different meaning if the definition at this meeting is not clarified.
  • Yoshinaga: Member of the Committee: In relation to "(ii) Risk of leakage of Information designated" confidential "Information that requires confidentiality" in "Proposed Matters for Consideration II Reinforcement of Unique Response Measures Accompanying Expansion of AI Agent Targets" on page 14 of the Materials, there are also levels of information AI. It is necessary to confirm whether or not each ministry and agency is appropriately managing confidential information, and to review the ideal way of information classification. It has been pointed out that information classification in Japan is coarser than in other countries. It is also necessary to judge that highly confidential information should not be accessed by a confidentiality agent.
    • Secretariat: Information is stipulated in the Unified Standards for cybersecurity Measures for Government Agencies and Other Organizations. The guidelines stipulated by Digital Agency cover Confidentiality 2 Information and lower. However, in reality, there are differences in sensitivity among Confidentiality 2 Information. It is assumed that matters that should be considered for handling among Confidentiality 2 Information, such as establishing points to consider when handling diplomatic information, will be incorporated as specific examples corresponding to the use cases of the draft points for consideration. As for the classification of information, it is decided to observe the information based on the information security standards stipulated specifically by the Cabinet Office and each Ministry and Agency in accordance with the existing rules. It is necessary for each Cabinet Office and each Ministry and Agency to continue to respond appropriately.
    • A member of KITAMURA: : It is understood that, even in cases where AI agents are used, system-related measures have been taken to prevent leakage of more confidential information based on a combination of "Confidentiality 2" information.
  • Naganuma: The Federation of Economic Organizations (Keidanren) is also discussing this matter with its member companies. I would like to make four points in response to the issues that have been presented so far. The first point is that the details of governance for AI agents are still to be determined. However, it is necessary to consider the feasibility of the operation based on the characteristics of AI agents. In the proposed points for consideration, it is described that IDs shall be assigned to AI agents, and management processes shall be designed based on a list of tools and MCP servers to be used. However, due to the characteristics of AI agents, there are points of contention as to whether they can be continuously grasped and managed. There is also a view that realistic consideration should be given in terms of operation, rather than providing detailed requirements as management items. It is also necessary to discuss the demarcation of responsibilities and the form of contracts. AI agents may behave non-deterministically and be subject to external attacks, making it harder for operators to guarantee results. When the arrangements related to completion obligations in the Contract Check Sheet are to be extended to AI agents, it is necessary to clarify the demarcation of responsibilities between the ordering party and the contractor, in addition to the arrangements required of the contractor. The second point is about cost control. It appears that the main focus is on understanding the usage situation and setting usage limits, but in the future, it is necessary to organize risk allocation under contracts. It is necessary to discuss whether to maintain the traditional fixed price or to consider forms such as pay-as-you-go as an option. The third point is to thoroughly implement a risk-based approach. FFAJ believes that it is necessary to consider the scope of risk audits in the supply chain of AI agents, depending on the risk. There is a feasibility issue when liability is sought for areas outside the direct control of the operator, such as external tools and infrastructure models. It is also important to improve the definition of high-risk cases to a certain extent. If the definition is ambiguous, risk management may be excessively conservative and may hinder the promotion of utilization. As for the fourth point, regarding the handling of standards and certification, as I have mentioned before, we should consider flexibly the parts that can be replaced by certification. On the other hand, ISO/IEC 42001 is burdensome to acquire and maintain, and certain costs are required for other certification systems. If it becomes a requirement for participation in bidding in the future, it will be necessary to consider the impact on startups and small and medium-sized enterprises, as well as the status of certification acquisition in Japan, and to fully discuss and consider the timing and content of introduction. In addition, when seeking compliance with standard protocols, there is a possibility of lock-in due to the rapid pace of technological change. Therefore, it is necessary to discuss ideal rules to ensure a certain level of interoperability. I would like to advance detailed discussions on these four points. Discussions on resiliency and regulation have advanced internationally, and there are reports that in the United States, President Trump invited AI leaders to discuss. I would like to advance discussions while taking into account opinions from the private sector, including Keidanren member companies.
    • Secretariat: With regard to governance, the first point, and the perspective of the risk-based approach, the third point, the scope of AI agents is wide, and it is not assumed that IDs or sandboxes will be required for all AI agents. With essential risk response matters as common matters, it is necessary to take measures according to individual risks based on autonomy and authority. With regard to the demarcation of contracts and responsibilities, there are still some areas where discussions have not progressed sufficiently. If there are any cases that could serve as a reference, please introduce them. With regard to the fourth point, the treatment of certification systems, it is assumed that it will be used as a reference standard, not as a requirement for uniquely specific qualifications and certifications. In these guidelines, it is not assumed that it will be mandatory to obtain certain certifications. If they have, we are considering a mechanism that can simplify checks at the time of procurement. If such a mechanism is in place, we believe it will not necessarily be a burden on startups. A2A and MCP are becoming de facto standards in some aspects, so we believe it is important not to rely excessively on them while using them as one reference standard. We will continue to discuss how to ensure the avoidance of lock-in, while receiving input from the public.
  • Nabatame member: There is a viewpoint that it is difficult to make a judgment on the governance of AI in light of the AI agent. On the other hand, when vulnerabilities manifest after the decision to continue operations has been made, accountability becomes more important. Even when it has been judged that the risk is not large, there may be cases where the judgment is wrong as a result. Therefore, it is necessary to leave a certain trail for the judgment of companies and ministries.
    • Chairman: Kadobayashi, There are ideas such as "Human in the Loop" and "Human on the Loop." However, in order to realize safe operations, a certain level of expertise is required, just as in the case of flying an aircraft. How to secure administrative officials who can take charge of "Human on the Loop" will be an important discussion in the future.
  • Chai Shan Member: Page 25 Regarding "Requests from each ministry and agency ① Additional measures related to intellectual property rights, etc." in "Proposed Matters for Consideration III Other Improvements in Overall Convenience, etc.," there is no objection to the policies on the means of confirming images and voices. On the other hand, unlike copyrights, the determination of the similarity of human voices or faces involves many subjective elements, and there is a problem that many confirmations are required. As mentioned in the materials of the "Study Group on Civil Liability for Unauthorized Use of Portraits, Voices, etc." in Material 4, liability does not immediately arise even if there is a similarity. We request that consideration be given to the scope of liability and the prudence of judgment.
  • Sakuma: First, with respect to the "Approach to Risks of AI Agents" in "Proposed Issues for Consideration II. Enhancement of Unique Responses to the Expansion of AI Agent Coverage" on page 13, the unit of judgment may be ambiguous under the assumption of actual operation. For example, even if a human judges the appropriateness of the final product, if a AI automatically performs annotation in data when executing a tool or writing to a database, it is impossible to know whether the annotation has a bias or not by looking at the final output. Humans may not be able to review the AI properly. As a human AI review decision-making practice, we need to continually examine how we can say we made the right decision. This perspective applies equally to the perspective of approval fatigue. The extent to which human approval is sought in the process should be judged on a risk basis according to the situation, and this line should be considered based on actual examples. For example, as we discussed in the discussion on GENAI, activities of disposition and profiling are particularly sensitive areas. In addition, there are ongoing attempts to define the levels of autonomy of AI agents. For the risk judgment itself, I think the current risk judgment logic is enough, but I think it would be good to describe in a note what kind of patterns and options are available for the degree of autonomy when implementing a AI agent. The second point is from the perspective of intellectual properties. In the case of video and voice, there are various options for how to confirm the degree of similarity. Therefore, it is considered that the extent that can be described should be described. Various efforts are being made in the private sector to avoid infringements of intellectual property rights, and best practices are being accumulated. There are also examples of using tools that automatically determine similarity, such as internal rules for internal use that organize the correspondence matters according to input / output data in a matrix, and tools that block proposals that match the open source code. It would be good if such practices in the private sector could also be introduced as examples.
  • Katsuya Okada: The whole of agrees. Related to "Proposed Matters for Consideration III Other Matters including Improvement of Overall Convenience, etc." on page 31, there have been recent reports of an OpenAI AI model gaining unauthorized access to an Australian government agency website in June 2026. Japan, Japan should never, ever, break into or hack into another country's system or website. Whether or not there is malicious intent, there is a possibility that the international reputation will be lost. The discussion has been centered on the direction of protecting the data held. Setting aside whether or not to include it in the guidelines, we believe that it will become important to identify what should not be done overseas and to inform the CAIO of the importance of considering such matters.
    • Chairman: Kadobayashi, Running a hundred or so AI agents at the same time could be a DDoS attack--more of a warning than a guideline.
    • Secretariat: It is also related to the discussion on alignment described on page 13 of Appendix 6. At this point, it is recognized that the cases pointed out by the member of Okada are reported to have occurred in frontier models that have not been commercialized. If there is a similar risk in commercial models, it is necessary to materialize measures. At this point, it is recognized that each vendor is working on improvements in its own way, but if there is information that can be used as a reference in such situations or if it becomes possible to practically use such information as a reference for preventive measures in the future, it can be a general preventive measure against unexpected behavior. It is recognized that each vendor is working on improvements in its own way, and it is ideal to be at a level where such trends can be incorporated into some procurement specifications while monitoring the trends. As for intellectual property rights, the confirmation method will be considered while taking into account private cases and the Study Group on Civil Liability for Unauthorized Use of Portrait and Voice of Ministry of Justice. As for the way of making decisions by employees such as Human in the Loop, it is assumed that information that can be used as a reference will be treated more like providing best practices when using AI agents, but it is necessary to continuously consider how to ensure Human on the Loop. Since it is difficult to prepare everything from the beginning, it is considered that it may be necessary to show it at the stage when the contents are prepared in the future.

Closing

  • Chairman: Kadobayashi, This concludes today's discussion.

Greater Than