The Next Question for AI Summits: How Do We Know AI Is Trustworthy?

Bringing the Hiroshima AI Process into practice through AI assurance

In its address to the United Nations General Assembly on 22 September 2026, New York time, Japan expressed its intention to host an AI summit at the earliest opportunity. It placed that ambition alongside the Hiroshima AI Process and efforts to co-create a safe, secure and trustworthy AI ecosystem with partners, including countries in the Global South.1

Against this backdrop, we believe one practical question deserves particular attention:

Who can use evidence that AI is trustworthy — and for which decisions?

The proposal is not to begin safety evaluation from scratch. It is to make its findings useful to businesses considering deployment and insurers considering the risk — and to give effective investment in safeguards appropriate recognition in doing business. Our proposal is to develop AI assurance into economic infrastructure that supports those connections.

1. Connecting the Hiroshima AI Process to real-world decisions

The Hiroshima AI Process is more than a statement of principles. Its International Code of Conduct for organizations developing advanced AI systems addresses lifecycle risk assessment and mitigation, internal and independent external testing, and updated technical documentation.2

The reporting framework launched by the OECD in February 2025 also provides a mechanism for organizations to report policies and practices aligned with the Code. These are substantive steps toward transparency and accountability.3

We want to build on that foundation at the level of specific business decisions.

Consider a manufacturer introducing AI into equipment operation. Knowing that a model has been evaluated is useful, but the business must also establish whether the findings apply to its equipment and operating conditions. Procurement teams, operational managers and insurers make different decisions and assume different responsibilities.

Reporting, technical evaluation and operational decisions should be connected through evidence — not collapsed into a single pass label. That is the starting point for our proposal.

2. AI assurance is not another declaration of safety

Recognizing that AI agents can affect external systems and physical environments, Japan AISI added “observation and control” to its evaluation-perspectives guide in July 2026. The update addresses autonomous behavior and interaction with the external environment.4

AI assurance is already an international field of work. UK government guidance describes measuring, evaluating and communicating AI trustworthiness, and emphasizes the need to demonstrate risk management in ways that trading partners in other jurisdictions can understand.5

Within that broader field, our focus is on enabling the party making a decision to establish what has been demonstrated, what remains unknown, and which operational decision the findings can support.

In logistics, for example, a business could go beyond describing its safety policy to examining whether processing was held when required information was missing, or whether execution followed the necessary approval. Findings should not then be carried over indiscriminately to another shipment or a changed system.

This is neither a guarantee against every AI incident nor a claim that one technology can conclusively prove safety. It is a way to make existing testing, auditing and risk management more useful to the decisions that follow.

3. AI exclusions expose a coverage problem that performance alone cannot solve

Insurance is a concrete example of why these connections matter.

In March 2026, US insurer HSB announced AI liability coverage addressing certain AI-related losses excluded by some General Liability policies. AI exclusions and coverage gaps are therefore already explicit product-design issues. This does not mean that AI losses are universally excluded across countries or insurance products.6

Separately, an April 2026 report from Japan’s Financial Services Agency and Ministry of Economy, Trade and Industry identifies insufficient historical loss data as a difficulty in evaluating and underwriting emerging risks, including AI and IoT.7

The first question is not simply whether premiums can fall. It is what information is missing when a business seeks coverage, and what could help an insurer reach a decision.

Evidence about controls and operations could help address part of that information gap. The proposal is not to replace loss experience with mathematics or automatically calculate unknown AI risk. It is to complement actuarial analysis and existing risk assessments with evidence relevant to the operation being considered.

Technical findings do not remove an exclusion from an existing policy. Coverage decisions and contractual arrangements remain separate. An assessment may also reveal weaknesses that justify further controls or a decision not to underwrite.

That is why the task should not sit with technology providers alone. Businesses deploying AI, assessors, insurers and reinsurers should examine together what evidence is useful for the decisions each must make.

4. Building a market that recognizes investment in safeguards

Our first two articles explored the connection between assurance and underwriting, and the economic incentives it could create. The wider ambition is a cycle:

Implement controls → establish their effectiveness and limits → use the findings in insurance, procurement and contracts → sustain investment in controls.

A business procuring AI might consider not only performance and price, but also whether the supplier can substantiate relevant operational safeguards. An insurer could evaluate that evidence against the exposure and use it when considering terms. If these connections work in practice, assurance becomes more than an additional engineering expense: it supports transactions and business continuity.

The Japanese report’s annex also discusses using Total Cost of Risk to assess cost-effectiveness and providing corporate risk analysis to insurers to help optimize coverage and premiums.7

We want to bring that perspective to AI operations. Better evidence is not automatically lower loss or better terms. The effectiveness of controls and the cost of verification must be assessed, rather than assumed.

Economic infrastructure does not mean making one product mandatory. It means enabling appropriate safeguards and supporting evidence to inform decisions repeatedly across organizations. Insurance is one important connection — not a substitute for safety or legal compliance, and not a prerequisite for every use of AI.

5. Why this matters for industrial resilience and economic security

Japan’s July 2026 AI Basic Plan treats controls, management and institutions as part of AI implementation capability. It also addresses excessive dependence on particular countries or companies, interoperability, and operational capability in strategically important domains.8

We see an industrial role for AI assurance in that context.

Businesses using AI in important operations should be able to judge the conditions under which it can be used, rather than rely solely on a provider’s account. They should be able to reassess changes and, when necessary, stop, switch or recover. The evidence and expertise needed for those decisions should also develop on the user’s side.

The objective is not to keep AI running regardless of circumstances. It is to stop what needs to be stopped while maintaining essential operations, including through alternatives.

Developing that capability alongside insurance and commercial arrangements could support industrial resilience and, in turn, economic security. That is our proposed contribution — not a claim that the government has designated AI assurance as national infrastructure.

For the same reason, an assurance ecosystem should not lock trust into one vendor’s product or one country’s models. Different technologies and assessors should be able to work together, while responsibility for decisions remains clear.

6. Three implementation questions for the emerging AI summit conversation

We propose three areas for practical international work.

First, develop a shared understanding of evidence for decisions. Developers, users, assessors and insurers should be able to understand what was evaluated, the conditions under which the findings apply, and what remains unverified. This is not a proposal for a universal certificate of safety. It is about connecting existing reporting and evaluation to distinct decisions, without requiring unrestricted disclosure of personal data or trade secrets.

Second, test the connection to insurance and procurement in defined workflows. A manufacturing or logistics pilot could examine what assurance findings add to underwriting or deployment review, whether their value justifies the effort, and how they remain useful after operational changes. Success should not be measured only by the number of businesses that receive a pass. The contribution to decisions — and its limits — should be made explicit.

Third, co-develop the capacity to use assurance with Global South partners. Japan’s plan includes cooperation on language and cultural diversity, skills and capacity building.8 Building on that direction, we propose working with businesses, assessors and insurers in regions such as ASEAN to design approaches suited to local operations, institutions and costs. Cooperation should leave decision-making capability with local partners, rather than simply deliver a system designed elsewhere.

Insurance and procurement incentives should not become new barriers for smaller businesses or emerging economies. Proportionate requirements and more than one viable way to provide the necessary evidence should be part of the design.

7. The connection GhostDrift aims to build

GhostDrift Mathematical Institute aims to connect technical findings about AI with evidence that can support real decisions. We have filed a related application in Japan: Japanese Patent Application №2026–210256.

We welcome joint work with AI-using businesses, assessors and auditors, insurers, reinsurers and insurance brokers, starting with a defined operational use case. Details of the underlying technology would be addressed in individual discussions under confidentiality arrangements.

As Japan moves toward hosting an AI summit, we believe the discussion should include both how justified trust can be established and how the resulting evidence can be used in real economic decisions.

From trustworthy AI to AI that businesses can choose with justified confidence — and a market that recognizes investment in using it responsibly over time.

That is how we propose to build on the international cooperation established through the Hiroshima AI Process: by developing AI assurance from a safety capability into economic infrastructure.

Primary sources and public materials

1 Ministry of Foreign Affairs of Japan, address by Prime Minister TAKAICHI Sanae to the 81st Session of the United Nations General Assembly, 22 September 2026, New York time.Relevant passage: the AI discussion in section 3, covering the Hiroshima AI Process, cooperation with the Global South, and the intention to host an AI summit in Japan.Official provisional English translation / Japanese original

2 G7, “Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems,” 30 October 2023.Relevant passages: actions 1–3, on risk assessment and mitigation, testing, documentation and post-deployment practices.Official text (PDF) / Publication page

3 OECD, “Launch of the Hiroshima AI Process (HAIP) Reporting Framework,” 7 February 2025.Official explanation of organizational reporting aligned with the Code and monitoring its voluntary adoption.Official announcement

4 Japan AISI, announcement of the Guide to Evaluation Perspectives on AI Safety, version 1.20, 7 July 2026.Relevant passage: the addition of observation and control for AI agent systems. Japanese-language source; descriptions here are summaries, not quotations from an official English edition.Official announcement

5 UK Department for Science, Innovation and Technology, “Introduction to AI assurance,” 12 February 2024.Relevant passages: sections 3 and 4.1, on assurance, decision-making and international interoperability.Official guidance

6 HSB, “HSB Introduces AI Liability Insurance for Small Businesses,” 18 March 2026.The announcement addresses AI-related exclusions in some General Liability policies. It described delivery through partner insurers subject to regulatory approval; it is not evidence of universal availability or coverage.Official announcement

7 Japan’s Financial Services Agency and Ministry of Economy, Trade and Industry, report of the study group on enhancing corporate risk management, 17 April 2026.Relevant passages: section III-2(3) and annex section 2(1), on emerging-risk data, Total Cost of Risk, and information provided to insurers. Japanese publication; the title is rendered descriptively in English.Full report (PDF)

8 Government of Japan, AI Basic Plan, Cabinet decision of 14 July 2026.Relevant passages: chapter 1 and chapter 3, section 3, on implementation capability, dependence, interoperability, and cooperation and capacity building with the Global South. Japanese original.Full plan (PDF)

Sources checked 24 September 2026. The implementation agenda and economic-infrastructure proposal are GhostDrift Mathematical Institute’s own views, not an agreed AI summit agenda, government or institutional endorsement, or an announcement of a partnership. The patent application is pending. This article does not announce an insurance product launch or promise coverage or discounts.