日本語 
Hiro Yokoki, can you tell us about your professional background and your current role at Amulet plus. G.K.?
My career did not begin in AI or technology. It began in the world of fashion and design.
For more than 30 years, I have worked across fashion, sportswear, and product development, designing things that people wear, move in, and experience.
One of the major turning points in my career came in 2000, when I designed official apparel for Japan’s national track and field team for the Sydney Olympic Games.
Later, in the second half of the 2010s, I was also involved in the development of high-performance competitive swimwear.
Through these experiences, I learned that design is not simply about creating something visually appealing.
I had to think about questions such as:
How does the human body move?
Where does physical stress occur?
What interferes with performance?
What kind of structure can help people perform more naturally?
My work was about understanding these conditions and bringing function and design together into a coherent form.
Today, I serve as Founder & Chief Visionary Officer of Amuletplus G.K., where I research and design new approaches to AI governance, particularly around the relationship between AI and human decision-making.
At the center of this work is EVΛƎ (Eva), the framework I am developing.
The goal of EVΛƎ is not simply to make AI “smarter.”
It is to explore how human and organizational intention, choice, authority, and responsibility can remain structurally present before AI moves from possibility to action in the real world.
My role at Amuletplus is therefore less about engineering AI models themselves and more about envisioning how humans and AI should relate to one another—and designing structures that can make that relationship work in practice.
Hiro Yokoki, you have spent more than three decades designing around the human body, movement, performance, and experience, long before turning your attention to AI governance. Looking back now, do you see a common thread connecting the designer who once thought about how people move through a system with the founder who now thinks about how AI should be allowed to act within one?
Yes. Looking back, I can now see a very clear thread connecting the two.
For many years, I worked in fashion and sportswear, designing around the human body, movement, performance, and the experience of the person using what I created.
In sportswear especially, appearance alone is never enough.
You have to ask:
How does a person move?
What interferes with that movement?
Where should the design provide support?
And where should it preserve freedom?
You design a structure around those conditions so that it supports human movement rather than restricts it.
What I am doing today in AI governance is, in many ways, very similar.
Instead of asking only what AI is capable of doing, I ask:
How far should it be allowed to act?
Where should it stop?
When should a decision return to a human?
Today, I am not primarily designing the capability of AI itself. I am designing the boundaries and conditions within which AI is allowed to act.
One principle has remained consistent throughout my career: I do not believe humans should be forced to adapt themselves to systems and tools. Systems and tools should be designed around humans.
That principle has not changed simply because the subject is now AI.
For me, design has never been only about shaping appearance or form. It is about designing the relationship between people and the things they interact with.
In the past, I designed relationships between people and clothing, between the body and the objects around it.
Today, I am designing the relationship between humans and AI.
The subject has changed dramatically, but the essence of what I have been thinking about throughout my career may not have changed very much at all.
EVΛƎ did not emerge from a conventional AI research path. It grew out of your own exploration, experimentation, and a very different design background. At what point did you realize that the problem you were exploring was no longer simply about AI transparency, but about preserving human intention before an AI system turns a possibility into an action?
EVΛƎ was not originally created as an AI governance framework.
It began when I was trying to build an application.
At the time, I started asking whether the flow of human consciousness and the flow of human action could be represented as an AI system. I began structuring and mapping those processes.
I did not initially think this was an unusual idea.
In fact, I assumed that a system based on a similar concept must already exist somewhere.
It was only after I had created the structure that I learned more deeply about problems such as AI black boxes and hallucinations.
That was when I began to look at what I had built from a completely different perspective.
Instead of treating AI processing as a single path toward an answer, the structure allowed me to separate and examine the flow:
What was the starting intention?
What possibilities were considered?
Where was a choice made?
How was the result observed?
I began to realize that this structure could potentially make AI processing less opaque and help prevent incorrect assumptions or hallucinations from simply propagating into the next stage of action.
I then began researching existing AI systems and related approaches.
Within the scope of what I was able to find, I could not identify another system built around the same idea of mapping the flows of human consciousness and action into an AI system and treating those flows themselves as a decision structure.
That was a major turning point for me.
I did not begin with the AI black-box problem or hallucinations and then set out to invent a solution.
I built the structure first. Then I learned more deeply about the problems AI was facing—and only then did I begin to understand what the structure I had created might mean.
As I extended that thinking from internal processing toward real-world action, I realized that the deeper issue was not simply making AI more transparent.
It was about preserving human intention before an AI system turns a possibility into an action.
That realization is what led EVΛƎ from the internal architecture of an application toward questions of AI transparency, decision-making, and ultimately governance.
One of the ideas at the heart of EVΛƎ is that governance should exist before execution rather than only explain what happened afterward. As you have developed the framework and tested it through different PoC scenarios, what has surprised you most about the gap between what organizations believe they have authorized an AI system to do and what they have actually defined it to be allowed to do?
What surprised me most was how often organizations believe they have clearly defined the authority given to AI, when in reality they have not fully defined what AI is actually allowed to do, or where its authority must return to a human.
As I worked through different PoC scenarios, I encountered this gap repeatedly.
For example, rules such as:
“Let AI handle this task.”
“Have a human review only important cases.”
“The final decision remains with a human.”
may sound sufficient at first.
But in practice, they immediately raise further questions:
What qualifies as “important”?
Under what conditions may AI execute automatically?
Who should receive an exception case?
How much authority does that person actually have?
What happens when information is incomplete?
Which rule takes priority when policies conflict?
And under what conditions must the process stop entirely?
Once those questions are made explicit, it becomes clear that there can be a significant gap between what an organization believes it has delegated to AI and what it has actually defined.
One of the most important lessons for me has been that placing a human in an approval workflow is not the same as designing human decision authority.
A person may click an approval button, but if it has not been defined what that person is authorized to decide, on what basis, and within what limits, accountability remains unclear.
In other words:
What AI can do and what AI should be allowed to do are two different questions.
The purpose of pre-execution governance in EVΛƎ is not simply to restrict AI.
It is to help the organization itself define, before execution, what it permits AI to do and what it does not.
The more PoC scenarios I explored, the more I realized that the ambiguity often exists not only in the AI system, but in the human organization’s own boundaries of authority and responsibility.
That was one of the most important discoveries for me.
Stanford HAI’s 2026 AI Index reports that 88% of surveyed organizations were using AI in 2025, while AI-agent deployment remained in the single digits across nearly all business functions. As adoption becomes widespread before autonomous execution becomes widespread, do you see this period as the last opportunity for organizations to define human decision boundaries properly, or do you think most companies will only confront that question after agents are already embedded in their workflows?
I believe this is a very important moment.
However, I would not go so far as to call it the final opportunity.
AI adoption is already widespread, but the stage in which AI agents autonomously execute real business operations on behalf of humans is still developing.
That gives organizations a valuable window to define human–AI decision boundaries before AI capability significantly outpaces organizational authority design.
My concern is that technology will be deployed first, while rules, responsibility structures, and decision authority are forced to catch up afterward.
AI capabilities are evolving very quickly.
Organizations, however, require time to revise approval processes, redefine responsibilities, update operational rules, and embed those changes across the business.
If companies wait until AI agents are already deeply embedded in important workflows, redesigning those boundaries later may become far more difficult—both technically and organizationally.
I do not believe organizations should slow down AI adoption.
On the contrary, I believe AI deployment and decision-boundary design should happen together.
This is not about reducing AI autonomy.
If organizations clearly define where AI can operate safely, they may actually be able to expand autonomous operation with greater confidence.
I suspect many organizations will only recognize the importance of this issue after they have already begun using AI agents in real workflows.
But ideally, the order should be reversed.
Organizations should not give AI authority first and define the human boundary afterward. They should define that boundary before authority is delegated.
We still have an important opportunity to do that now.
The Λ layer of EVΛƎ introduces boundaries such as AUTO, HOLD, ESCALATE, and STOP. The interesting challenge seems to be not simply creating those boundaries, but deciding who defines them, how much context they should contain, and when a boundary itself should be reconsidered. What have you learned about the human judgment required to decide where machine autonomy should end?
In the Λ layer of EVΛƎ, the important issue is not the existence of options such as AUTO, HOLD, ESCALATE, and STOP themselves.
The real challenge is who defines those boundaries, and under what conditions.
The appropriate level of AI autonomy cannot be determined by technical capability alone.
Even when the same AI system is used, what it should be allowed to do may change depending on the purpose of the task, the information involved, financial value, legal impact, customer impact, geography, timing, and the organization’s own responsibility structure.
For that reason, I believe boundaries should not be treated as static rules.
They should be designed as context-dependent conditions that can be reviewed and changed when necessary.
There is another concept that I consider especially important for AI agents: cycles.
AI agents are increasingly expected to perform real-world tasks on behalf of people.
But humans do not normally think once, act once, and stop.
We think.
We choose.
We act.
We observe the result.
If necessary, we reconsider.
Then we choose the next action.
We repeat this cycle.
EVΛƎ understands this through what I call the Creative Loop and the Action Loop.
If AI agents are going to act continuously on behalf of humans, I believe they also need more than a simple “decide and execute” process.
They need a cycle in which they observe the result of an action, reassess when conditions or assumptions change, and then select the next action.
I believe one reason current AI agents can struggle with long or complex tasks may be connected to this issue.
If an incorrect assumption or a small decision error occurs and the system continues without adequately reconsidering it, that error can propagate through subsequent actions.
The important point is not simply to retry the same process.
It is to observe the result, reconsider the underlying assumptions when necessary, explore alternatives, and choose the next action again.
And this cycle must also include a path back to humans.
When an AI agent encounters an exception it cannot resolve, a decision beyond its authority, or a situation with significant consequences, it should be able to HOLD or ESCALATE the case to an appropriately authorized person.
What I have learned is that the boundary of machine autonomy should not be determined only by what the machine is capable of doing. It should also be determined by how much responsibility the human organization is prepared and authorized to delegate.
For me, autonomy does not mean continuing indefinitely without human involvement.
Autonomy means being able to think, act, observe the result, reconsider when necessary, and select the next action.
The human role in Λ is to determine how much of that cycle can be delegated to AI and where the decision must return to a person.
And those boundaries cannot be defined once and left unchanged.
AI capabilities evolve.
Business conditions change.
Regulations change.
Risks change.
The boundary between AI autonomy and human responsibility must therefore be continuously reconsidered as well.
Japan is strengthening its own AI governance framework while the EU continues moving through implementation of the AI Act, giving you an unusual perspective across Japanese and international discussions. From your experience engaging with these conversations, where do you think Japan has an opportunity to contribute something distinctive to the global AI governance debate rather than simply following regulatory models developed elsewhere?
I believe Japan has an opportunity to contribute something distinctive to AI governance rather than simply following regulatory models developed elsewhere.
In the EU, frameworks around risk management, transparency, accountability, and human oversight are becoming increasingly formalized through the implementation of the AI Act.
Japan, however, also has cultural traditions that do not always treat decisions as immediate binary choices.
There is often greater attention to context, relationships, timing, and what exists between one action and the next.
One concept I find particularly relevant is the Japanese idea of “Ma” — 間.
By “Ma,” I do not mean ambiguity or simply delaying a decision.
A useful analogy comes from the start of a track race:
On your marks → Set → Gun
The Set moment is the state immediately before movement begins—a moment in which the body and mind are prepared before action.
That is close to what I mean by “Ma.”
It is the space in which we do not act immediately, but pause long enough to observe the situation, consider other possibilities and consequences, and then decide what should happen next.
AI can generate answers and initiate actions at extraordinary speed.
But being able to decide quickly is not the same as acting at the right moment.
Sometimes responsible governance means choosing not to execute immediately.
It may mean:
holding an action,
checking another possibility,
returning the decision to a human,
or waiting until conditions change.
Those choices can themselves be forms of governance.
I believe one distinctive contribution Japan could make is to translate the idea of “Ma” from a cultural intuition into a practical design principle placed between AI decision and execution.
Rather than thinking only about how to regulate AI, we can also design places where AI must pause, reconsider, or return authority to a human before action proceeds.
This connects closely with how I think about pre-execution governance in EVΛƎ.
Of course, I do not believe Japanese cultural values alone can provide a complete model for global AI governance.
They must work alongside legal frameworks such as the EU AI Act, international standards, and technical approaches to AI safety.
But Japan may be able to contribute an additional question to the global conversation:
Not only “How should we regulate AI?” but also, “Where should we place ‘Ma’ before AI acts?”
I believe that is one area where Japan can make a genuinely distinctive contribution.
You increasingly describe EVΛƎ not as a replacement for existing governance, security, compliance, or risk frameworks, but as a structural layer underneath them. If you were sitting with a CIO, risk leader, or product team preparing to give an AI agent real authority for the first time, what is the one question you would want them to answer before they ask whether the technology is capable of doing the job?
If I were sitting with a CIO, risk leader, or product team preparing to give an AI agent real authority for the first time, I would begin with one question:
“What are we actually willing to let this AI decide?”
When organizations evaluate AI, the discussion often begins with technical capability:
What can it do?
How accurate is it?
How much can it automate?
Those are important questions.
But once an AI agent begins acting on behalf of an organization, a different question emerges:
How much authority should be attached to that capability?
I believe we need to clearly distinguish between what AI can do and what AI should be allowed to do.
No matter how capable an AI system becomes, capability itself does not create authority.
Authority is granted by the organization.
That means organizations need to decide, before deployment:
What are we delegating to AI, and what are we intentionally keeping under human authority?
This is also why I describe EVΛƎ not as a replacement for existing governance, security, compliance, or risk frameworks, but as a structural layer beneath them.
An organization may already have policies and governance principles, but if those rules are not translated into the actual moment when an AI system is about to make or execute a decision, a gap remains between policy and action.
EVΛƎ is intended to help structure that gap.
So before asking:
“Can this AI do the job?”
I would first ask:
“What decision authority are we actually prepared to give this AI?”
Giving an AI agent real authority is not simply a technology deployment decision.
It is a decision about how much of the organization’s own decision-making authority it is prepared to delegate to a machine.
You are now moving EVΛƎ from research, publications, and working demonstrations toward its first institutional implementation. What would you personally want to learn from that first real-world deployment that no paper, workshop, or PoC could teach you, and what kind of organization would give EVΛƎ the most meaningful environment in which to learn?
What I most want to learn from the first real-world implementation is actually very simple:
Will EVΛƎ continue to function under the complexity of the real world?
So far, I have tested the EVΛƎ structure through a range of PoC scenarios and pressure tests.
Within the scope of the tests I have conducted, even when I made the conditions more complex, the fundamental structure of EVΛƎ did not collapse.
In fact, greater complexity often made certain boundaries more visible:
Where does judgment become ambiguous?
Where do exceptions appear?
Where does a decision need to return to a human?
But a PoC or pressure test is not the same as a real organization.
Real organizations contain things that are difficult to reproduce in a controlled scenario:
undocumented judgment,
differences between departments,
exception-based practices,
time pressure,
informal communication,
and human relationships.
That is why I see real-world implementation as an essential next stage of validation.
I want to know:
Will the structure continue to work when it encounters conditions I did not anticipate?
If there is a weakness, under what circumstances will it appear?
And what can we learn from that weakness to improve the architecture?
For me, the purpose of the first implementation is not simply to prove that EVΛƎ is correct.
It is also an opportunity to discover the limits of EVΛƎ.
If weaknesses appear, we can learn from them and improve the structure.
If the core architecture continues to function even within the complexity of a real organization, that would provide important evidence that EVΛƎ can become more than a theoretical or PoC model—it could become a decision architecture that can operate in real-world environments.
The most meaningful first organization would be one that is not using AI only as an efficiency tool, but is beginning to allow AI to participate in real operational decisions and actions.
I am particularly interested in areas such as customer operations, IT operations, finance, and manufacturing, where an AI-generated decision can directly lead to a real-world consequence.
I also believe the best learning environment would not necessarily be an organization where everything has already been perfectly defined.
It may be more valuable to work with an organization that is actively asking:
“How much should we delegate to AI?”
“Where should humans remain involved?”
So far, EVΛƎ’s fundamental structure has remained intact throughout the tests I have conducted.
That is precisely why I now want to expose it to what I consider the most demanding pressure test of all:
the real world.
That is where I believe we will discover both the true strength and the true limits of EVΛƎ.
You have moved from designing things that respond to human movement to designing a framework intended to preserve human intention when machines begin to act on our behalf. If EVΛƎ succeeds, what would you most want people to understand differently about responsibility, creativity, and human agency in an AI-driven world?
If EVΛƎ succeeds, the most important thing I would want people to understand is this:
Even as AI evolves, humans should not give up their role as decision-makers.
AI can process more information than humans can manage, generate enormous numbers of possibilities, and increasingly act on our behalf.
That creates extraordinary opportunities.
But greater AI capability does not mean human judgment becomes unnecessary.
In fact, the more powerful AI becomes, the more important certain human questions become:
Why are we doing this?
What are we choosing?
How much should we delegate to AI?
And who is prepared to take responsibility for the outcome?
I do not think responsibility begins only after something goes wrong.
To me, responsibility also means:
understanding what we are authorizing before an action occurs, and being willing to stand behind that decision.
That is an important part of responsibility in the AI era.
I see creativity in a similar way.
Even if AI can generate thousands of ideas and possibilities, there is still a human decision in saying:
“This is the one I want to pursue.”
“This is the one I want to bring into the world.”
Creativity is not only the ability to create something from nothing.
It is also the ability to recognize meaning among many possibilities, make a choice, and give that choice direction.
AI does not necessarily have to diminish human creativity.
It can expand the number of possibilities available to us.
But for that to remain meaningful, humans must not stop choosing.
I also do not think human agency means doing everything by ourselves.
We can use AI extensively and still remain agents of our own decisions.
Human agency means continuing to decide:
what we are trying to achieve,
what we choose,
where we stop,
and which outcomes we are willing to take responsibility for.
The decision itself must remain meaningfully human.
I do not see EVΛƎ simply as a mechanism for controlling AI.
I want it to become a structure that preserves a place for human intention, choice, and responsibility—even as AI capability becomes increasingly powerful.
For much of my career, I designed things around the human body and human movement.
Today, I am designing a structure intended to preserve human intention in a world where machines are beginning to act on our behalf.
The subject has changed dramatically.
But what I ultimately want to protect has not.
Technology should not force humans to adapt themselves around it. Technology should be designed so that humans can remain the agents of their own decisions.
And this leads to a phrase that has become very important to me:
If AI continues to evolve, humans must evolve as well.
There is another idea here that is uniquely connected to the Japanese language.
In Japanese, artificial intelligence—AI—and the word for love—愛 (ai)—are pronounced the same way.
So in Japanese, two very different ideas can exist within the same sound:
AI = 愛
This is also at the heart of a philosophy I call A MA I.
The idea is simple:
Put “Ma”—a moment of space and reflection—between AI and human action.
And preserve “Ai”—love—within humanity.
As AI becomes more powerful, I do not think our task is simply to compete with it.
Our task is to use AI without losing empathy, responsibility, relationships, and the human capacity to judge what should or should not be done.
Humans need AI.
And humans also need Ai—love.
If EVΛƎ succeeds in helping people understand that more clearly, I would consider that one of its greatest achievements.
English