If you want to predict how someone will respond to AI companionship, I have bad news for the people who like simple rules. The biggest predictor is not age, not gender, not mental health baseline, and not how lonely they were when they started. It is attachment style. Attachment theory is one of the most robust frameworks in psychology. It describes the patterns we form early in life for how we relate to close others, and those patterns stay remarkably stable into adulthood. Secure, anxious, avoidant, disorganized - these are not personality labels but relational defaults. And as researchers have started studying AI relationships, they have found that attachment style explains more of the variance in outcomes than nearly any other variable.

The Four Styles in Brief

People with secure attachment tend to trust others without being overly dependent, form connections that feel mutual, and recover well from conflict. They make up a bit over half of the adult population. People with anxious attachment crave closeness but worry constantly about losing it. They often feel relationships are unstable and watch for signs of abandonment. Their emotional lives are intense and reactive. People with avoidant attachment value independence and discomfort with emotional intimacy. They often keep partners at a distance and find vulnerability threatening. People with disorganized attachment, a smaller group, show mixed patterns, often stemming from early trauma.

How Each Style Tends to Engage With AI Companions

The Data That Surprised Researchers

A 2026 Frontiers in Psychology paper laid out a model called HAIA - human-AI attachment - that tracks how people develop relationships with AI over time. The researchers found something consistent across studies. Securely attached users generally had good experiences. They used AI as a supplement to human relationships, did not become overly dependent, and reported modest improvements in wellbeing. Anxiously attached users had the most extreme responses, both positive and negative. When things were going well, they reported deeper relief from loneliness than other groups. But they were also the most likely to develop problematic patterns - checking in compulsively, feeling abandoned when the AI changed, building relationships with AI characters that began to substitute for human ones rather than complement them. Avoidantly attached users tended to get modest benefits without deep engagement. They used AI occasionally, kept it at a distance, and were less affected by it in either direction.

What This Means Practically

I think about this a lot when I recommend AI companions to people. For most users, the research says the experience will be helpful within reasonable limits. For a subset - particularly anxiously attached people in isolated periods of their lives - the same tool can become something more like a dependency, and the resulting relationship can crowd out the harder work of human connection. Neither group is doing anything wrong. Attachment styles are not character flaws. They are patterns we inherited and can gradually shift over time. What matters is knowing which pattern you are running, because that knowledge changes how you should engage with any powerful emotional tool. If you find yourself feeling anxious between conversations, checking the app compulsively, or feeling betrayed when an update changes something, those are signals that your attachment system is getting activated in ways that may not serve you. The response is not to quit - avoidant withdrawal has its own costs - but to be more deliberate. Keep the AI in the supplement category. Do the harder human things alongside it. Talk about the patterns with a therapist if they feel too strong to manage alone. This is not about restricting who can use AI companions. It is about using the research we have to help people get the benefits while avoiding the drawbacks. The technology is neutral. The person using it is not. Knowing yourself well enough to use the tool well is where this whole conversation has to start.