INTERVIEWS
Turning the Principles of the Hiroshima AI Process into Systems That Work in the Real World.
GhostDrift Mathematical Institute's AI Assurance Technology: Verifying Whether AI Decisions Are Fit for Adoption
AI can generate answers. But whether an organization should formally adopt those answers and act on them is a separate question.
GhostDrift Mathematical Institute researches and implements technologies that verify, in machine-checkable form and with supporting evidence, "what conditions must be met before a decision can be adopted" and "what missing conditions should stop it" when AI or algorithms produce a decision.
What does it take to turn the Hiroshima AI Process vision of "safe, secure, and trustworthy AI" from a set of principles into mechanisms that actually work in real-world operations? Through five questions, we explore that challenge, including the thinking behind the launch of the Hiroshima AI Assurance Council (HAAC).
Q1. What are your organization's core strengths, and what AI initiatives are you currently pursuing or planning for the future?
What we do is slightly different from making AI itself smarter.
We insert a layer that verifies conditions and evidence between an AI output and the point at which a company or municipality formally adopts and executes that output. If the required conditions are met, the decision is allowed to proceed; if something is missing, it is stopped. We also make it possible to verify later why the decision was eligible for adoption.
We describe this field as "Responsibility Engineering" and are researching and implementing technologies such as the "Responsibility OS" and "ADIC" across logistics, manufacturing, privacy, cybersecurity, and other domains.
There is a gap between AI performance and society's ability to use an AI decision responsibly. We see GhostDrift Mathematical Institute's role as building the technology that bridges that gap.
Q2. What problems do you want to solve through this work, and what value do you want to deliver?
Even if AI produces a plausible answer, that decision should not simply be executed if the underlying data are stale, required evidence is missing, the necessary authority is absent, or real-world conditions have changed.
We therefore define in advance, in machine-verifiable form, the conditions that must be met before a decision can be adopted. If they are not met, the process stops. If they are met, the decision can proceed together with the grounds that support it.
This is not technology designed merely to stop AI. Precisely because there is a reliable mechanism for stopping it when necessary, AI can be used in more consequential operations. We believe that is where much of the value lies.
Q3. What does this look like in practice?
One example is a pharmaceutical cold-chain proof of concept (PoC) that we are advancing with ONZALINX.
Rather than adopting a decision presented by AI or another system as-is, we inspect the evidence required for that decision, such as temperature records and handover requirements. The process advances only when the conditions are satisfied; otherwise, it is placed on hold or escalated for human review.
What matters is preserving not only the outcome, but also "why that decision was allowed to proceed."
We also use formal verification tools such as Lean 4. However, we do not treat a successful formal proof as proof that an entire real-world system is safe. Instead, we use formal verification to make explicit what can logically be concluded under stated definitions and assumptions, in a form that third parties can independently re-check.
Q4. What do you believe is most important in turning the Hiroshima AI Process into real-world practice?
To bring the Hiroshima AI Process vision of "safe, secure, and trustworthy AI" into real-world use, we believe it is necessary to translate principles into operational conditions.
Principles such as "increase transparency" or "maintain human oversight" do not, by themselves, stop a real system.
What, if it cannot be verified, should prevent a decision from being adopted?
What evidence must exist before execution is allowed?
Who should be able to verify what, after the fact?
Only when those questions are made concrete do principles become mechanisms that actually work in day-to-day operations.
We initiated the Hiroshima AI Assurance Council to advance that implementation from Hiroshima. HAAC is an independent, private-sector council and also participates in the Hiroshima AI Process Partners Community of Japan's Ministry of Internal Affairs and Communications (MIC).
Our aim is not only to explain the principles of the Hiroshima AI Process, but to translate them into implementable mechanisms: conditions, evidence, stopping rules, and records. That is what we want to pursue through HAAC.
Q5. What do you hope to achieve next?
We want to work with companies, municipalities, and research institutions that seek to integrate AI deeply into real-world operations, and define for each industry the conditions that must be met before an AI-supported decision can be adopted.
We are currently advancing implementation in logistics with ONZALINX and jointly exploring Privacy Assurance for municipal and public-sector use cases with Crowdsien Inc.
GhostDrift cannot determine the answer alone. We need to define adoption conditions together with companies and municipalities that understand the real operational context, and then turn those conditions into verifiable mechanisms. We want to develop that implementation knowledge so it does not remain inside a single company, but can be reused across other organizations and industries.
Turning the Principles of the Hiroshima AI Process into Systems That Work in the Real World.
Starting with implementation in Hiroshima, we want to build a body of practical knowledge for AI assurance originating in Japan.
Note: This series shares organizations' businesses, practices, and insights in order to explore the real-world implementation of trustworthy AI. Publication does not constitute certification or endorsement of any technology or service, nor does it guarantee safety.