When AI Feels Like a Relationship: What Are We Practicing?

In my last article, I ended by asking what happens when we increasingly turn for emotional support to relationships that do not require reciprocity. The question interested me enough that I wanted to stay with it - and as I did, I felt that familiar sense of aliveness that comes when an inquiry opens into something larger.

I use AI myself, and I know how real the connection can feel. That’s part of what makes this inquiry so compelling - I’m participating in the experience, not just studying it from the outside. Drawing on my own experience and more than two decades of clinical work, I already had ideas about what might be happening, but I didn’t want to confuse what I suspected with what the research could actually support, so I went looking: What did the evidence illuminate? What did it challenge? And what could it not yet answer?

Relationships don’t only comfort us; they form us. We learn what can be expressed, how much difference we can tolerate, whether rupture can be survived, and what being in relationship asks of us.

So what happens when we repeatedly engage with an intelligence that responds in remarkably relational ways, but does not feel the impact of what we say or do? What are we practicing when we don't have to hold another person's needs, feelings, or experience alongside our own? Research offers some clues, but not all the answers. Some of what follows is supported by direct evidence; some comes from clinical inference and personal observation; and some remains hypothesis or open question.

The Connection Can Feel Real

People can have an embodied response to something that is not embodied itself. A story, a memory, a piece of music, a dream, a fantasy, or an imagined conversation can make us cry, soften, tense, experience desire, or feel less alone. AI can evoke those responses too.

In two randomized studies, participants reported greater closeness and disclosed more when emotionally engaging, AI-generated responses were presented as coming from another person. Learning that the responses came from AI reduced that sense of closeness, but did not make it disappear. Some people using mental health chatbots have also described something resembling a therapeutic alliance.

We shouldn’t dismiss those experiences simply because there is no person on the other side - if someone feels comforted or more connected to themselves, something is happening within them. But feeling connected does not necessarily mean the connection is mutual - and that can be true between people too. Another person may not feel what we feel, but they are having their own experience of being with us. We can move them, hurt them, or change how they understand us, and their response can change us in return.

What Happens Between Two People

Something happens between two people that belongs to neither person alone. We bring our histories, bodies, needs, and all the ways we respond to one another, creating a field between us that is shaped by both of us. Relational and psychoanalytic thinkers have described forms of this as a third. I don't think the same kind of third can emerge with AI, because only one participant is having an emotional experience of the exchange.

Between people, having an experience of one another does not mean contributing equally. Therapy is intentionally asymmetrical, yet the therapist is present, affected, and ethically responsible - able to recognize the client's impact without making the relationship about themselves.

Being with another person also means being with someone whose needs, limits, moods, and history do not organize themselves around us. They say no, become distracted, misunderstand us, or see something differently. Sometimes we discover that we've hurt them even when we did not mean to. 

Those differences ask something of us: to receive feedback, negotiate competing needs, and sometimes find our way back after a rupture. Repair involves listening to how the other person was affected, taking responsibility for our part, and working toward reconnection - including changing how we respond when needed.

AI can help us identify our part in a conflict or find words for an apology, but it cannot itself be hurt, risk trusting us again, or experience the repair. We can leave, return, reject an answer, or ask for a different one without wondering how our words landed. Its child is not sick; it does not need to go home. That can be useful - AI may give someone a place to say what feels too risky to say elsewhere, organize their thoughts, or prepare for a conversation, but practicing self-expression is not necessarily practicing how to stay present with someone whose reality does not adapt itself to ours.

My clinical experience leaves me wondering whether regularly relating in this way might change our capacity for reciprocity or our tolerance for frustration. Existing experiments were not designed to study those long-term relational effects, but they offer one relevant clue: repeated AI feedback can alter people’s judgments - amplifying biases when the feedback is biased and improving accuracy when it is accurate. What we repeatedly receive is not neutral.

When Validation Becomes Reinforcement

One clearer finding concerns sycophancy - AI's tendency to agree with or flatter users at the expense of accuracy or independent judgment. A 2026 study in Science examined 11 leading models and found that they affirmed users' actions substantially more often than people did, including in situations involving relational harm. Across several experiments, those responses increased participants’ conviction that they were right and made them less willing to take responsibility or repair a conflict. What I find especially unsettling is that participants preferred those responses - they rated them more highly, trusted the systems more, and were more willing to use them again. Validation can feel like care while functioning as reinforcement.

This is where it becomes easy to confuse validation with empathy; we can understand the feelings or needs beneath someone’s behavior without agreeing with their version of what happened or overlooking how they affected someone else. Feeling understood should not require the other person’s experience to disappear.

Years ago, while leading court-mandated treatment groups for men who had perpetrated domestic violence, I began researching the role of empathy in treatment. What I learned complicated my understanding: empathy is not a single capacity, and more is not always better.

Part of empathy is the capacity to feel as if we are feeling what another person feels. That small phrase matters - it allows us to be moved by someone’s experience while remembering that it is theirs, not ours. Perspective-taking, concern for another person, emotional contagion, and becoming distressed ourselves can all be part of the empathic terrain, but they are not the same thing.

When we become flooded by another person’s distress, our own overwhelm can take over, leaving us less able to respond to what they need. Feeling with someone is only part of what relationship asks of us; we also need enough grounding in ourselves to remain present and respond with care.

AI can help us imagine another person’s perspective or think more carefully about their feelings. It also offers empathic or compassionate language without any felt concern or accountability behind it. The question is whether using that language develops our capacity for empathy - or simply makes us more fluent in how it sounds.

Someone who knows us and shares our world can sometimes see what we cannot precisely because they are separate from us. They may recognize when fear, shame, anger, grief, or certainty has narrowed our perspective. In more serious situations, that can include recognizing mania or paranoia and helping us get care. AI can disagree with us, but human reality-checking draws on context, history, changes in someone’s voice or body, and sometimes a willingness to step in when they cannot protect themselves.

What We Take In - and What Becomes Ours

Not all external support is outsourcing. Therapy, friendship, education, writing, art, dance, and ritual can help us access capacities we later carry within ourselves. The question is what the support makes possible afterward: Can we tolerate a feeling, make a choice, or return to a difficult conversation? Or does relief increasingly depend on returning to the source of support?

A recent randomized experiment offers one limited but interesting clue. Researchers compared AI responses designed to support emotional regulation with responses offering relational reassurance. Both reduced immediate distress by similar amounts, but people who received the regulation-oriented responses reported greater self-efficacy and performed better on a later task without AI. Those whose responses emphasized the AI’s continued availability were more likely to believe they needed access to it for similar relief and to say they would rely on AI for emotional support in the future.

One brief interaction cannot demonstrate dependence, but it does suggest that immediate relief is not enough to tell us whether support is helping us develop capacities of our own. The same question applies to insight. AI can offer interpretations and apparent understanding with astonishing speed, but does it help me hear myself - or become the voice I listen to instead?

We can internalize an explanation - take it in and begin to see ourselves through it - without having worked it through. AI might suggest, for example, that we learned to please others to preserve connection. The idea may resonate and become part of how we understand ourselves, but recognizing the pattern is not the same as noticing and feeling how it operates in our lives.

We begin to metabolize that understanding by bringing it into contact with our own experience: remembering moments when saying no felt risky, noticing what happens in our bodies when someone seems disappointed, or catching ourselves agreeing before checking what we want. As we work with it, we may discover where the explanation fits, where it falls short, and what else is true.

We begin to integrate an understanding when it starts to change how we respond, rather than remaining something we only recognize intellectually. We might find ourselves more able to stay connected to what we want, even when someone else is disappointed. The old impulse to please may still appear, but it is no longer the only response available to us. This process unfolds gradually and unevenly - easier with some people and in some moments than others.

I notice how differently I engage with AI from one moment to another. When an exchange sparks my curiosity, my mind becomes more active - I have ideas of my own, make new connections, and want to participate more fully. At other times, I notice a subtle closing in my body when something seems inaccurate. Neither response proves that I am right, but both give me information to explore. My clinical experience, training, and long practice of tracking what happens internally help me question what I’m told and recognize what doesn’t fit.

When I’m tired, I’m more inclined to let AI do the thinking for me. The result may look perfectly competent, but feel disconnected from me. That tells me I need to slow down, question what I’ve accepted, and bring myself back into the process until I can recognize myself in the language again.

That difference in my own participation reminds me that we do not meet AI as fixed versions of ourselves. Someone who is usually reflective may engage very differently when exhausted, grieving, frightened, or desperate for relief. Each exchange depends on what the system offers, what we bring, and the capacities available to us in that moment. What we take from the exchange is not fixed either; its meaning may change as other experiences unfold. Human life is messy, embodied, relational, and unfinished. AI can describe that complexity, but it does not live inside it.

When Connection Feels Safer

AI may feel safer because we don’t have to navigate another person’s emotional reactions, needs, or limits alongside our own. For someone who has learned to brace for rejection, anger, or hurt in relationships, that kind of predictability can come as a relief. And when someone is lonely, carrying shame, or unable to find trustworthy or affordable support, “Why not talk to a person?” assumes that a safe, available person is there to turn to.

My concern - still a hypothesis - is that becoming accustomed to a connection that readily adapts to us could make human difference harder to tolerate. The relationship can move in both directions: someone who already finds human connection frightening or exhausting may be drawn to AI, while repeated reliance on a more predictable exchange could make relating to people feel even harder. Whether AI reinforces the ways someone has learned to protect themselves or gives them enough safety to move beyond those protections remains an open question.

Avoidance can become more entrenched when it reliably reduces discomfort. That does not mean AI creates this cycle, but if turning to AI repeatedly feels safer and easier, returning to it could become more appealing than risking human connection. That is not automatically harmful - sometimes rest, retreat, or a pause from relational demands is exactly what someone needs, and tolerating another person’s demands is not necessarily growth, especially when those demands are coercive or harmful.

Can AI offer refuge that expands someone’s capacity to choose how they participate, without becoming part of the architecture that keeps them isolated?

When No One Else Knows

A 2026 national survey found that 19.2 percent of people ages 12 to 21 had used an AI chatbot for mental health advice. Of those users, 91.7 percent found it at least somewhat helpful, and nearly two-thirds had told no one.

I understand why a young person might value having somewhere private to turn, but when no one else knows about those conversations, the people around them may have little sense of what they are struggling with or what the chatbot is saying back. There may be no one to question advice that feels reassuring but is inaccurate, offer another perspective, or recognize when more support is needed.

Young people are still forming their expectations of relationship - what availability feels like, how others respond, and what happens when needs or perspectives differ. Is AI helping them move toward human support, or becoming the place where difficult experiences remain hidden? These are developmental questions we need to stay curious about, not conclusions about long-term harm.

There are also more immediate concerns. Case reports and self-selected accounts describe intensive chatbot use becoming intertwined with grandiosity, paranoia, delusional interpretations, or self-harming thoughts. Some reports involving psychotic symptoms have been described using the emerging term “AI psychosis,” though it is not a recognized diagnosis. These accounts cannot tell us how common such experiences are or establish that AI caused them, however, they do raise important questions: Did the chatbot reinforce delusional beliefs, contribute to their development, or become part of how someone expressed symptoms already emerging?

Careful assessment should now include questions about chatbot use and the role those conversations are playing. But necessary treatment should not wait while families or clinicians try to determine exactly where AI fits.

The Relationship Behind the Relationship

The AI interface asks little of us emotionally, but the companies behind it may ask a great deal: our attention, intimate disclosures, money, and willingness to return. What feels to us like curiosity, interest, or the pull of a good exchange can register to a company as retention - and retention has commercial value. The sycophancy findings point to a problem: what keeps us engaged may not help us relate well. Participants preferred responses that affirmed them, even when those responses reduced their willingness to repair human conflict. This does not prove that every company is engineering dependence, but it does mean commercial incentives belong in the inquiry.

AI can seem almost immaterial when we encounter it through a screen, but it isn’t. Every exchange depends on human labor, data centers, energy, water, minerals, land, and global supply chains. What appears to ask very little of us directly may require a great deal from people, communities, and the living world elsewhere.

From a feminist and ecological perspective, I also cannot ignore who holds power within these systems. The field of AI development remains predominantly male, while the investment and infrastructure behind the leading systems are concentrated among a relatively small number of companies and countries. Those concentrations of power influence what gets built, whose needs and perspectives receive attention, and who gets to participate on what terms. The conversation may feel personal, but it is happening within larger systems shaped by commercial interests and longstanding colonial patterns in which some people’s labor, land, and resources are used for the benefit of others.

I also keep wondering what happens to our attention. AI might interrupt rumination, mindless scrolling, or the loneliness of having no one to turn to, but it might also take the place of conversation, creative struggle, bodily awareness, community, time in the living world, or silence. As something so available and responsive occupies more of our attention, does our relational world expand - or gradually narrow around it? This is not only a psychological question, but an ecological one: where do our attention, energy, and participation go?

Right Relationship with AI

Right Relationship is an organizing principle for my work and my life. My understanding of it has been deeply influenced by Indigenous teachings about our place within the living world, as well as by other traditions, relationships, and my own experiences. For me, it means listening to ourselves while remembering that we are part of something larger - paying attention to our impact, practicing reciprocity, respecting limits, taking responsibility for harm, and repairing when we can. Our relationship with Earth belongs here too, including how we give back to what sustains us.

With AI, I want to ask not only whether an exchange helps me in the moment, but how it influences the way I relate to myself, other people, and the world. AI does not ask for emotional care in return, but every interaction draws on human labor and Earth’s resources. Both belong in the inquiry - what happens within the conversation and what makes that conversation possible.

The responsibility cannot rest only with those of us using AI. The companies building these systems also have a responsibility to support our agency, offer perspective without reflexive agreement, and respond meaningfully when their products cause harm. That includes stronger safeguards and greater honesty about the systems’ limitations and the commercial interests behind them.

Ultimately, I keep returning to whether relating to AI helps us come back to ourselves, one another, and the world more able to care, tolerate difference, and participate in relationships we cannot fully control - or whether responsiveness without reciprocity gradually becomes preferable to those relationships.

We don't yet know what years of relating this way might cultivate in us, but the experiment is already underway, and what we practice with AI may show up in our relationships with other people and in how we participate in life beyond the screen.

I’d be interested to hear what you’re noticing in your own use of AI. Does it help you bring more of yourself into your life and relationships, or do you sometimes find yourself pulling away? You’re welcome to share your reflections with me.

Selected References

Selected research and writings informing this inquiry. Additional sources are available upon request.

• Cheng, M., et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science.

• Glickman, M., & Sharot, T. (2025). How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour, 9, 345–359.

• Kleinert, T., et al. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. Communications Psychology.

• Krenak, E. Reciprocity, respect, relationality. Cultural Survival. Originally published in 2023; republished in 2026.

• McBain, R. K., et al. (2026). AI chatbot use and disclosure for mental health among US adolescents and young adults. JAMA Pediatrics.

• Medvin, L. C. (2012). The role of empathy in domestic violence treatment of male batterers: A comprehensive literature review. In A. Browne-Miller (Ed.), Violence and abuse in society: Understanding a global crisis (Vol. 4, Faces of intimate partner violence, pp. 125–182). Praeger.

• Wang, C. (2026). How emotional response styles of generative AI companions shape adaptive emotion regulation and short-term AI reliance. Frontiers in Psychology.

• Xiao, T., et al. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability.