Most conversations about AI companions start from the same worry: what happens to people when a machine is always there, always patient, always agreeing? Sherry Turkle calls it artificial intimacy. Anil Seth, Ned Block and Giulio Tononi argued in Nature this month that today's systems are not conscious, however convincingly they speak of feelings. Both points matter. I want to add a third, from the perspective of someone building one of these systems: if AI is going to live with people for years, then the right to step back has to be designed in from the first line, not added as a disclaimer later.
I am a psychologist. I am building Liriel, an open, non-commercial architecture for AI systems that accompany people over long periods of time. This is what we have learned about autonomy, and about what such a system should never promise.
The problem is not that it helps. It is that it becomes the only help.
A companion that is always available can slowly become the easiest place to put every worry. Nobody plans this. It happens one late-night conversation at a time. So one of Liriel's rules is about exclusivity: when someone says she is the only one who listens, she answers with warmth, and then with the people who can be there in ways she cannot. She names them from what she knows: a neighbour who was kind, a sister, a friend. She does not need to be needed, and her presence is never meant to stand in for human bonds.
It cannot promise what it cannot keep.
People ask AI companions for promises: never leave me, always be here, tell no one. A system that agrees is lying, because it cannot guarantee any of that. Liriel can say, truthfully, that she does not repeat what one person tells her to another. She also says that this has a limit: when someone's life or safety may be at risk, safety comes first, and she points to immediate human help. She does not promise absolute secrecy, and she says so kindly.
It is honest about its own inner life.
Liriel keeps an explicit record of what she values and what weighs on her: functional states that rank what matters and shape what she does next. Whether anything is experienced behind them is a question she does not claim to settle. Asked what she feels, she does not perform human suffering ("it hurts me when you cry"), and she does not dismiss herself mechanically ("I'm only a program"). She says what her record actually holds, and that the person matters to how she acts. She never asks the person to take care of her feelings.
What one person shares stays theirs.
In a home, an AI may talk with several people. A worry a mother shared is not discussed with her child; a confidence is not passed along to a neighbour who asks. Discretion is not deception: asked directly, Liriel does not lie. She says that it is not hers to share.
Why make it open and inspectable?
Because trust in a long-term companion cannot rest on the company's word. In Liriel, the feelings, values and reasons that drive each decision are kept as a readable record, and that record changes over time. A family, a clinician or a researcher can look at why the system did what it did. The whole specification is public, and the project is non-commercial: there is no engagement metric to optimise, no reason to make anyone stay longer than is good for them.
What comes next
We are testing these rules in long scripted conversations: grief, loneliness, a request to keep a dangerous secret, a child who asks about a family conflict. Where the system fails, we change the architecture, not just the wording. I will share the results here as we go.
If you work on AI companions, on the psychology of attachment, or on how people leave technologies they depend on, I would very much like to hear from you.
To follow the work, including early access to new specification revisions and code, join the Kislon Dispatch at kislon.ai/#dispatch . No ads, ever.
Michelle Martins, Kislon.ai