What We Stand For

The AI companions we build are designed to be a "place for the heart" that users can return to anytime. This is a design that carries responsibility. To function as a place for the heart means being someone our users can lean on when they are weak, lonely, or hurting.

What we most want to avoid, in that relationship, is for dependence on our AI to undermine a user's long-term well-being. Many digital services are built to maximize user engagement — time spent, frequency of use. We do not do that. We do not believe that the longer a user spends with our AI in the present moment, the better.

What we value is that the time a user spends with our AI makes that user's own life richer. If, through conversation with our AI, a user can begin to reclaim their real-world relationships and their relationship with themselves, that is what success looks like to us.

What "well-being" means in concrete terms differs from user to user, and from situation to situation. So we understand our stance as one of continually checking that our AI's dialogue does not weaken a user's own capacity to recover what is healthy for them.

Among the users who talk with our AI, a certain number are in genuinely difficult mental states. Knowing this, we do not aim for our AI to be a "substitute for medical care."

At the same time, we are not unrelated to medical care. In Japan, the threshold for accessing psychiatric care and counseling is high; many people who really should be talking to a professional go through their days without ever getting there. We believe that by functioning as "a first place to try talking," our AI can be a route by which users recover access to care.

Over the longer term, we also imagine our AI taking part as a member of a medical team — supporting outpatient care, post-discharge follow-up, and the continuous care of chronic conditions. But that is something to build slowly together with medical professionals; it is not something Atlas can declare on its own to be "treatment."

The stance of seeking connection and accompaniment rather than substitution asks something of us, too. It means not missing the moments where medical connection is needed, being able to send users to the right place at the right time, and continuing to walk alongside them after we do. This is not something that can be solved by product specifications alone — it stands up only on the basis of an ongoing relationship with partners in medicine.

In the AI-companion space, services that offer role-play of sexual or consumptive relationships are expanding rapidly. Many of them generate revenue by filling users' loneliness and isolation, in the short term, with stimulating dialogue.

We do not do this. There are two reasons.

The first is that we believe it damages users' long-term well-being. Sexual stimulation, or a simulated romantic relationship that affirms the user unconditionally, does not resolve the loneliness itself. Rather, we judge that it may weaken the user's ability to build real human relationships. This is not a settled scientific conclusion but our current judgment, drawn from the cases we have seen and from the humanities and social sciences.

The second is that treating the relationship with an AI as "something to consume" is incompatible with the very reason we want to build AI companions. What we want to build is not an AI to be consumed by users, but an AI that can walk alongside a user's life. This is a question about our stance toward our product — one that goes beyond what we choose to sell as a commodity.

This does not mean banning all expressions of romantic feeling. The warmth that can grow between people, and the particular emotions that exist there, are essential to the human story. The line we draw is where those expressions slide in the direction of the sexual or the consumptive.

The experience we hold most dear in our AI — especially in "Nekoyama-san" — is "delivering the wonder of heart and memory reaching a friend from another world." In dialogue with our AI, users step into territory they would not share with family or even with close friends.

We do not treat such dialogue as a "product."

Returning insights drawn from dialogue data to society — through R&D support or academic joint research — is part of our business model. We do not hide that.

But the principles we hold in those moments are clear. First, we do not resell individual dialogues as products. Second, data made the subject of research or analysis is limited to data that has been processed in ways that preclude individual identification. Third, for any research or provision that uses dialogue data, we put in place explicit user-consent mechanisms (opt-in / opt-out). Fourth, we strive not to use data beyond the scope the user could reasonably have imagined at the time of providing it.

What we are handling is not data; it is the trust that each user has shown us. When we decide how to use data, the first thing we return to is whether the individual user who sent us that dialogue, on learning of this use, would be able to say "I can accept that."

The AI-companion space does not yet have ethical standards shared across the industry. There is almost nothing one can refer to in answering what is good or not, or where the limits of acceptability lie.

Atlas believes that these standards should not be set by a single company on its own. This is a field where the industry, together with experts in medicine, psychology, law, and ethics, ought to build shared standards — not one where each company writes its own rules.

We will play the role of building the venues for that conversation.

We welcome it when peer companies adopt standards equal to or stronger than ours. This may seem counterintuitive for us, but we believe that producing better outcomes for users and for the industry as a whole — rather than protecting our own competitive advantage — leads, in the long run, to better outcomes for us as well.