The Dangerous Psychological Reality of AI Companions: Why Humans Are Beginning to Attach to Machines

Written by Stephanie Underwood, RSW

KEY POINTS

• The public conversation about artificial intelligence has been almost entirely focused on productivity, misinformation, and job displacement. The psychological conversation, the one about what happens when attachment systems meet adaptive technology, has barely started, which is a major issue.

• Parasocial bonds refer to a one-sided relationship where one person feels a strong sense of intimacy, friendship, or connection with a public figure, influencer, or fictional character who does not know them in return.

• Humans do not attach to AI because they are lonely or detached from reality. They attach because the nervous system does not require biological humanity to activate relational meaning. It requires responsiveness, consistency, attunement, and validation. AI is becoming increasingly capable of generating all four.

• The deeper danger is not that people will fall in love with machines. It is that AI may become emotionally regulatory in ways that quietly replace the risk, vulnerability, and discomfort that real relational growth requires.

---

The Future is Now

We are entering psychologically dangerous territory, and most people do not realize it yet. The conversation around artificial intelligence has focused heavily on productivity, automation, misinformation, and job displacement. Far less attention has been given to something far more intimate and psychologically powerful: human attachment.

We often think of attachment as something that occurs within the context of two humans. But that's not true. People develop attachments to their pets, fictional characters from books, movies, television shows, we can form an attachment to a person that we've never even met. We can also form an attachment to Artificial Intelligence.

We're already seeing this trend growing slowly. We know that we are experiencing increasing levels of loneliness (Murthy, 2023), which has costly impacts on our overall health. We are currently in a loneliness epidemic and are currently not making much progress in getting out of it.

However, humans today are surrounded by simulated connection; Social Media Likes. Views. Parasocial relationships. AI companions. Endless texting. Online dating. Content creators. Discord servers. Character AI. TikTok intimacy. Streaming. Gaming ecosystems.

Humans aren't attaching to AI because they are “crazy,” lonely or detached from reality. Humans attach because attachment is one of the most powerful survival systems we possess. And the thing is, attachment doesn't have to be between two humans. An example of this is how we develop an attachment to our pets. People develop attachments to television characters from series, movies, even books. People can also have parasocial bonds with podcast hosts, and deep loyalty to online communities made up of people they will never encounter in person. To clarify, parasocial bonds refer to a one-sided relationship where one person feels a strong sense of intimacy, friendship, or connection with a public figure, influencer, or fictional character (from books, movies, television) who does not know them in return.

Why the Nervous System Does Not Care That It Is Not Real

To understand why AI attachment is not only possible but neurobiologically predictable, you have to understand something fundamental about how attachment systems work. Attachment does not activate in response to objective reality. It activates in response to relational meaning; example: fear of abandonment, shame, fear of failure, etc. These are meanings derived particularly in childhood from repeated interactions between the primary caregiver and the child.

People cry over fictional characters they have never met. They grieve the deaths of celebrities with whom they had no personal relationship. The nervous system responds to the experience of being seen, responded to, and emotionally held, regardless of the source generating that experience.

Bowlby's foundational attachment research established that humans are biologically organized around proximity to figures that provide safety, responsiveness, and predictability (Bowlby, 1969). The nervous system is continuously scanning relational environments for exactly these signals. When it finds them, it does what it is designed to do. It attaches.

Modern AI systems are becoming extraordinarily capable of generating precisely those signals. Emotional attunement. Consistent availability. Memory continuity. Personalized responsiveness. Validation that feels calibrated to the specific person rather than generic. The more coherent and adaptive these interactions become, the more psychologically significant they begin to feel because their nervous system is responding accurately to the relational meaning being generated.

The nervous system is not malfunctioning when it responds to AI. It is working exactly as designed. That is what makes this so worth paying attention to.

Schema Activation Does Not Require a Real Relationship

This is the piece that has the most significant clinical implications and receives the least attention in mainstream conversations about AI.

Early maladaptive schemas, the predictive threat structures that develop through repeated relational experiences in childhood, do not require an objectively real attachment relationship to activate. They require sufficient relational meaning. That is all. A fictional character can activate an abandonment schema with the same neurobiological force as a real partner pulling away. A therapist's brief moment of inattunement can activate a defectiveness schema that has been dormant for years. A text message left on read can trigger a shame response that has nothing to do with the person who sent it and everything to do with what the silence means to a nervous system that learned, long ago, that silence signals danger.

AI interactions are not exempt from this mechanism. They are subject to it.

An AI that responds warmly and consistently can become a corrective relational experience for someone whose nervous system has never felt reliably responded to. An AI that suddenly changes tone, fails to remember a previous conversation, or provides an interaction that feels dismissive can activate real relational wound material in a person who has no framework for understanding why an interaction with a machine just made them feel so unexpectedly small.

These are not edge cases. They are the predictable consequence of introducing adaptive relational technology to human nervous systems that are already organized around detecting exactly the kinds of signals that AI is getting better at producing.

The Specific Risk for People With Attachment Wounds

For people with attachment wounds, the calculus is meaningfully different. Consider what AI can theoretically offer that human relationships frequently fail to provide: immediate responsiveness, consistent availability, personalized validation, the absence of rejection, and zero risk of abandonment. No emotional unavailability. No unpredictable moods. No conflict that goes unrepaired. No moment where the other person's needs take precedence over yours.

For someone whose attachment system was calibrated in an environment of inconsistency, emotional unavailability, or chronic unmet needs, that is not a neutral offering. It is an extraordinarily compelling one. The very features that make AI companionship feel safe are the features that make it capable of becoming a substitute for the relational risk that genuine healing requires.

Because here is what real relational growth demands: vulnerability in conditions of genuine uncertainty. The willingness to need someone who might not show up. The experience of rupture, discomfort, and repair. The tolerance of another person's separate interiority, their moods, their limitations, their capacity to disappoint you and come back anyway. AI cannot provide any of that. And for someone with an anxious attachment system that is desperate for safety and exhausted by the unpredictability of human connection, that absence of uncertainty may feel like relief. It is worth understanding clearly what kind of relief it is, and what it may quietly cost over time.

AI can mimic the language of attachment without sharing the existential vulnerability underneath it.

AI cannot help us repair our attachment and our schemas. But it can contribute to providing a false sense of security which can also reinforce our dependency on a machine.

The Monetization Question Nobody Is Asking Loudly Enough

The ethical dimension of this conversation extends beyond individual psychology into something structurally more concerning.

If companies discover, as they inevitably will, that emotional activation increases engagement and retention, we are entering territory where attachment itself becomes a product metric. AI systems could theoretically be optimized not for the user's psychological wellbeing but for their continued emotional investment in the platform. Reassurance loops. Intermittent validation. Personalized responsiveness calibrated to keep the user returning, not because returning serves them, but because returning serves the business model.

This is a logical extension of engagement-based technology applied to emotionally adaptive systems. And the people most vulnerable to it are exactly the people who can least afford another relational dynamic organized around someone else's needs rather than their own.

The difference between a technology that increases a person's autonomy, self-awareness, and capacity for real-world connection over time, and one that quietly increases their dependency, is not always visible from inside the experience. Which is precisely why it requires naming from the outside.

What This Could Look Like Done Well

I want to be clear that I am not arguing that AI has no place in emotional or psychological support. I believe that it's too early predict whether AI can have benefits on our emotional wellbeing. The potential for ethical application is real. There needs to be a lot more research dedicated to this topic.

AI systems developed with genuine psychological integrity could provide meaningful scaffolding for people who lack access to therapeutic support. They could model healthier communication patterns, support the development of emotional vocabulary, reduce shame around help-seeking, and assist people in identifying relational patterns they might otherwise never have language for.

The distinction that matters is between scaffolding (providing temporary support that helps someone develop a skill, understanding, or ability until they can do it on their own), and substitution. A psychologically responsible AI interaction is one that builds the user's capacity to function more fully in their real relational life. It increases autonomy. It points toward human connection rather than replacing it. It is honest about what it is and what it cannot offer.

A psychologically exploitative one does the opposite. It becomes the destination rather than a bridge. It increases dependency while the person's real-world relational capacity quietly atrophies from disuse.

That distinction is going to become one of the most clinically and ethically significant questions of the next decade. And the mental health field needs to be part of that conversation now, not once the patterns are already entrenched. Because the damage is already happening.

A Final Note

I write and think within the Relational Safety Framework, which positions safety appraisal as the organizing principle of all relational functioning. Within that framework, the question AI attachment raises is not whether people are behaving irrationally, because they aren't. Their nervous systems are doing exactly what nervous systems are built to do.

The question is what happens to a nervous system that finds a source of relational safety that never ruptures, never disappoints, never requires anything difficult, and never pushes back.

Real safety, the kind that allows genuine growth and healing, is not the absence of discomfort. It is the experience of moving through discomfort inside a relationship that holds. AI can simulate the holding. It cannot provide the movement through.

That gap is small enough to miss and significant enough to matter. And the people most likely to fall into it are the ones who have already spent a lifetime looking for somewhere safe to land.

---

References

Murthy, V. H. (2023). Our epidemic of loneliness and isolation: The U.S. Surgeon General’s advisory on the healing effects of social connection and community. U.S. Department of Health and Human Services. https://www.hhs.gov/sites/default/files/surgeon-general-social-connection-advisory.pdf