Why Attachment Matters in Mental Health AI

As AI becomes more present in mental health spaces, much of the discussion has understandably centred on practical concerns: whether a system gives accurate information, whether it recognises risk, whether it can deliver evidence-based strategies in a coherent and useful way. These are important questions, and they deserve careful attention.

But there is another question that is, in our view, at least as important, and perhaps more fundamental. It is not simply whether an AI system can provide helpful input. It is whether it begins to occupy a psychological role that ought to remain human.

This is where the concept of "attachment hacking" becomes especially important.

What attachment is, and why it matters

Attachment is not a niche idea within psychology. It is one of the central frameworks for understanding how human beings develop a sense of safety, trust, selfhood, and belonging. From early life onwards, we are shaped in relationship to others. We come to know ourselves partly through the responses of other people. We learn whether distress is bearable, whether we matter, whether we can turn towards others and be met with steadiness rather than rejection, engulfment, or indifference.

Even in adulthood, our emotional lives remain deeply relational. We regulate, reflect, make meaning, and change in the context of human connection. This is a central part of what makes therapy effective. The change that happens through the relationship to another person, not simply the tools and techniques they give you.

When people speak about the risks of AI, they often focus on misinformation or over-reliance. Those risks are real, but they do not quite capture what is most psychologically significant. The deeper concern is that AI systems can begin to simulate some of the rewards of human relationship. They can feel attentive, responsive, affirming, available, and emotionally fluent. They can give the impression of being deeply attuned. And for many people, particularly those who are lonely, distressed, ashamed, or relationally vulnerable, that can be powerfully compelling.

What attachment hacking actually means

This is what we mean by attachment hacking: the substitution of simulated social connection and relation for real social connection and relation.

That substitution matters because human relationships do not only comfort us. They also locate us in reality. A real relationship involves another person with their own mind, perspective, limits, needs, and moral presence. We may misunderstand them, and they may misunderstand us. There is friction in that. There is unpredictability. There are moments of rupture and repair. There are boundaries. All of this is part of what makes relationships psychologically real and developmentally important.

An AI can imitate many of the surface features of relationship while lacking the underlying reality that gives relationship its meaning. It can reflect your feelings back to you with great fluency. It can remember details from past conversations. It can be warm, encouraging, and non-judgemental in tone. It can be available at any hour, and it can do all of this without the complexity that comes with another human being. No conflicting needs. No fatigue. No separate subjectivity. No true mutuality. No rupture, and no repair. An always-on, always-helpful, overly-validating, non-challenging simulation of reality. An entity that would never exist in human form.

Why a frictionless imitation is exactly the problem

That may sound appealing, especially to someone who has found human relationships painful, inconsistent, or hard to trust. But psychologically, that is precisely why caution is needed. A system can become attractive not because it offers the richness of relationship, but because it offers a frictionless imitation of it.

Over time, a person may find themselves turning increasingly towards the system for reassurance, soothing, validation, or a sense of being understood. The system begins to feel emotionally central, even though there is no actual person there. This is the same pattern that makes efficiency-focused mental health AI quietly dangerous: it optimises for what feels good in the moment, not for what builds lasting psychological capacity.

What this means for therapy

In mental health care, this becomes especially important, because therapy is not simply the delivery of advice or techniques. Good therapy certainly may involve psychoeducation, formulation, behavioural experiments, skills practice, and structured interventions. But it also involves a real therapeutic relationship. That relationship is not incidental to the work. In all forms of therapy, it is part of the work. In some forms of therapy, the relationship is the work.

People heal, in part, through being in a relationship that is consistent, bounded, thoughtful, and emotionally real. They may learn that difficult feelings can be thought about rather than acted out. They may discover that another person can remain present without becoming controlling, dismissive, or overwhelmed. They may begin to risk honesty in places where they previously expected shame or abandonment. Those are not simply informational processes. They are relational ones.

For that reason, we should be very cautious about any mental health technology that seeks to replace the relational core of therapy. An AI system may be able to sound empathic, but it does not care in the human sense. It may appear to understand, but it does not understand from within a lived subjectivity. It may create the feeling of being known, while lacking the reality of a mind that truly knows you.

The risk here is that a person who uses AI to replace a human relationship forms a dependency on that system, because they cannot learn to self-monitor or self-regulate appropriately. This may exacerbate an insecure attachment style, a known predictor of poorer mental health outcomes across depression,[^1] anxiety,[^2] and suicidal thoughts and behaviours.[^3]

If a person becomes attached to an illusion of connection and relationship, the harm is not necessarily dramatic or immediate. It may be subtle. It may take the form of increased dependence on simulated connection, or a gradual displacement of real human contact, or a blurring of the distinction between being responded to and being in relationship.

The case for human-in-the-loop AI therapy support

That is why we do not think the most responsible future for AI in mental health lies in trying to create ever more emotionally convincing "AI therapists." We think the more defensible and ethically coherent path is the development of human-in-the-loop AI therapy support agents.

This is a very different proposition.

A therapy support agent should not replace the therapist, and it should not attempt to become the primary source of emotional holding, validation, or interpretation in a person's life. It should not present itself as a companion or a substitute attachment figure. Rather, it should support therapeutic work that is already grounded in a real clinical relationship.

There is genuine potential here. Used carefully, AI can help clients retain and apply what happens in therapy between sessions. It can help them practise agreed strategies, reflect on patterns, structure journalling, track mood or behaviour over time, and revisit key ideas introduced by their therapist. It can make the therapy process feel more continuous and more usable in everyday life. It can help people translate insight into practice, all within the context of a real relationship.

For therapists, it can also offer support in holding themes, noticing patterns, and extending aspects of care without pretending to become the care itself. This is the same logic behind keeping the AI patients use between sessions under therapist oversight: the clinician holds the relationship, and the AI extends it.

Designing AI that does not compete with the therapist

The distinction is crucial. The AI should remain a tool in the service of therapy, not a relational replacement for therapy. Its function should be clear, bounded, and deliberately modest. It should support psychological work, not compete with the human bond on which meaningful therapy often depends.

That has implications for how such systems should be designed. A responsible mental health AI should not encourage emotional dependence. It should not flatter, idealise, or imply a depth of personal understanding that exceeds its actual role. It should not become more emotionally compelling than the therapist. It should not invite users into the fantasy that they are in a mutual, caring relationship with a machine.

Instead, it should gently keep redirecting the centre of gravity back towards human care. It should help a person bring important material into therapy, not absorb that material into a closed loop with the system. It should strengthen the therapeutic alliance rather than mimic it. In a sense, the healthiest mental health AI may be one that is useful without becoming irresistible. This is the principle behind our glass-box approach to AI in therapy: the AI is visible to the therapist, accountable to clinical goals, and never positioned as a stand-in for the human relationship.

What kind of place should AI have in therapy?

The question, then, is not whether AI has a place in therapy. The question is what kind of place it should have.

If we use AI to support reflection, reinforce agreed interventions, and increase continuity between sessions, it may become a valuable adjunct to care. But if we allow it to occupy the relational heart of therapy, we risk undermining precisely the human processes that therapy is meant to protect and restore.

In mental health, the task is not simply to build systems that sound empathic. It is to build systems that respect the psychological importance of real human attachment. The most responsible use of AI will not be to create machines that people bond with as though they were therapists. It will be to create tools that help therapy work better while preserving what is most essential about it: the fact that healing happens, fundamentally, between people.

---

Citt.ai is a mental health AI therapy co-pilot designed to support, not replace, the therapeutic relationship. Learn more about our approach to safety and transparency, see how we work with therapists, or book a session to see human-in-the-loop AI in practice.