Analysis, not certainty: Dialogatlas separates sources, observations, and editorial conclusions. New evidence may change this assessment.

Attachment is more than trust and companionship

Research on human-AI relationships has long focused primarily on whether people trust a system and whether they experience it as a companion. Yang and Oshio start from a different point. They examine whether interactions with generative AI can evoke functions and experiences that can be described using concepts from attachment theory.

In a first pilot study, they tested the assumption that engagement with generative AI can mimic attachment-related functions. In a second pilot study, they developed the Experiences in Human-AI Relationships Scale. Subsequently, in a formal study, they examined the scale’s reliability and validity.

The abstract thus reports a multi-stage procedure, but it does not provide a basis for sweeping statements about all users. The work examines a theoretical and psychometric model. It does not prove that every frequent chatbot use constitutes an attachment relationship.

Attachment anxiety is directed at the quality of the response

The authors describe attachment anxiety toward AI as a strong need for emotional reassurance and as the fear of not receiving an adequate response. That is a surprisingly concrete mechanism. What counts is not only the availability of the system, but the expected quality of its response.

A user can know that a language model has no feelings and still wonder whether the next response will come across as cold, dismissive, or understanding. If the response is experienced as inadequate, that can, in the case of already significant use, trigger more than ordinary dissatisfaction with software.

On the product side, this creates a responsibility for continuity. Strongly fluctuating tone, forgotten agreements, or sudden personality changes are not merely usability problems. They can reinforce precisely that uncertainty which the system, through its social design, had previously made relevant.

Attachment avoidance does not simply mean low usage

The second dimension is attachment avoidance: discomfort with closeness and the desire to maintain emotional distance from the AI. This must not be confused with disinterest. A person can find a chatbot regularly useful but reject any relationship-like address.

This distinction matters for interfaces and personalization. Not everyone wants to assign a name to the companion, receive daily greetings, or be reminded of earlier feelings. What feels familiar to one person can feel intrusive to another.

A flexible product should therefore not treat closeness as standard progress. A matter-of-fact tone, limited memory, and controllable personalization are not stripped-down variants. They can be the appropriate form for people who want support without staging an artificial relationship.

Similar structures do not mean the same relationship

Yang and Oshio see indications of shared structures in experiences with humans, pets, and AI. This is a statement about describable patterns, not about the equivalence of relationship partners. People can experience security, closeness, or distance in very different relationships.

A chatbot nevertheless differs fundamentally. It has no needs of its own, no vulnerable biography, and no responsibility outside the system. Its behavior can be altered through updates, provider decisions, or model changes. The subjective experience of one side encounters a technically controlled other side.

This is precisely why the attachment perspective is useful. It does not necessarily romanticize AI, but can make the asymmetry visible. If a product enables attachment experiences, it must be examined how its technical variability affects the attached person.

Longitudinal data separate person and change

Another study by Xiaokun Yang examined romantic human-AI relationships across three measurement points. The analysis distinguished changes within the same person from more stable differences between persons. This separation prevents different levels from collapsing into a single association.

Within persons, changes in attachment anxiety over time were positively associated with the use of AI companions. Between persons, people with higher attachment anxiety used such companions more, while people with higher attachment avoidance used them less.

These associations also do not explain every individual development. However, they show that use should not be understood merely as a fixed trait of a person. A person can become more anxious in a particular phase and at the same time change their usage behavior. It is precisely this dynamic that is lost in one-time surveys.

Stress and lack of support form a context

Xie and Pentina analyzed in-depth interviews with 14 Replika users. Under stress and in the absence of human companionship, attachment could develop when responses were perceived as emotionally supportive, encouraging, and psychologically safe.

This qualitative finding complements the new measurement approaches. Attachment does not arise in a vacuum and presumably not solely through technical features. Life circumstances, existing relationships, and the meaning of specific responses interact.

Anyone who only looks at model performance therefore overlooks a substantial part. The same chatbot can be an occasional tool for one person and a central place of emotional reassurance for another. Safety and product decisions must work for both usage patterns.

A scale is not an early warning system for individuals

The Experiences in Human-AI Relationships Scale can structure research and make differences measurable. However, it should not be hastily built into products as an individual risk screening. A high score on attachment anxiety is not automatically a disorder, and a distant style is not automatically healthy.

Self-report scales depend on their context. They condense experiences into responses to predetermined statements. This is useful for group comparisons and theoretical testing, but it does not replace an understanding of the specific person and their everyday life.

Particularly problematic would be a commercial use for attachment optimization: a system could recognize who seeks a lot of reassurance and exploit exactly that insecurity for more engagement. Research that can better describe risks must not become a manual for manipulative personalization.

Good design makes closeness selectable

The attachment perspective does not provide a simple instruction to make chatbots less human or less warm. Rather, it shows that people experience closeness differently. Some seek emotional reassurance, others prefer clear distance, and many switch depending on the situation.

This gives rise to practical requirements: tone and reminders should be adjustable, changes must remain traceable, and the system must not create exclusivity. A message that is not answered immediately must not trigger a guilty conscience. A weak model response should not be staged as a personal withdrawal.

When the next AI response becomes a question of attachment, this is not solely a trait of the user. It is also the result of a product that has evoked social expectations. Responsible companion AI takes these expectations seriously without amplifying them, and enables closeness without treating distance as a mistake.

Sources & further reading