Attachment is the term psychologists use for the most significant relationships that hold people’s lives together.

But what makes an attachment relationship different from all others? According to the pioneering American-Canadian developmental psychologist Mary Ainsworth, it must meet the following six conditions.

(1) The relationship must persist over time, having (2) been formed with a particular subject who cannot be replaced. (3) The relationship must be emotionally significant, (4) involve the desire to remain close to them, and (5) cause distress at involuntary separation. And finally (6) when threatened, the person must seek security and comfort in their relationship with that subject.

The first five conditions describe any lasting affectionate bond, such as a close friendship. Only the sixth turns it into an attachment bond specifically. Simply liking someone, feeling close to them or spending hours together isn’t enough to create an attachment relationship.

Interestingly, though, nothing in these six conditions specifies a human on the other side.

This was already noticed long before the emergence of AI chatbots with the power to converse. Deities, pets and objects (ranging from children’s blankets and stuffed animals to cherished household items and even phones and computers) have all been found to supply security and comfort when someone is feeling distressed.

Today, AI companion apps are commonly sold on the promise of lasting and close relationships. Many of their users describe feeling real grief when an app is updated or withdrawn. But do such experiences meet the narrow definition of attachment?

The honest answer is: we don’t really know yet.

What the evidence suggests

Research on human-AI relationships has so far recorded users’ feelings about chatbots, their frequency of use, and their reactions to chatbot withdrawal. It hasn’t tested these relationships comprehensively against the six conditions described above.

Questionnaires that measure attachment to AI are being developed, validated and used, but most appear to take for granted that the relationship is an attachment. The language of attachment – and even “hyper-attachment” – has become standard in public coverage and in many academic papers too.

A recently published survey of 7,027 people in four countries (Germany, China, South Africa and the US) reported that at least a third of AI chatbot users show behaviour indicative of an attachment relationship. However, its three questions asked whether people consider the chatbot a friend, are polite to it, and miss it after a few days.

At most, these questions touched on emotional significance and mild separation anxiety. None asked whether people turn to chatbots for security and comfort when truly distressed.

Why does this matter?

Whether relationships between humans and AI chatbots count as attachment matters because the word “attachment” travels. It shapes how clinicians and parents read behaviour, how regulators frame consequences, and how companies describe their products.

Much of the evidence and public discussion on human-AI chatbot relationships rests on how often people use these apps, and how they feel about them. Yet, such measures cannot separate attachment from other processes.

People report feeling less lonely after chatting with an AI companion, which resembles seeking comfort. However, as some researchers have pointed out, the available studies didn’t assess whether emotions are actually regulated through the chatbot, and therefore conflate distress relief with attachment.

Some AI users react to losing an AI companion as they would to losing someone real, which matches distress at separation. But that condition applies to all affectionate bonds and doesn’t, on its own, identify attachment.

Calling these relationships attachments therefore goes further than the evidence currently allows. And likewise, the evidence against attachment isn’t conclusive.

In a small pilot study of human-AI relationships in China, for example, people’s attachment towards AI barely matched their attachment towards real people. Although this could suggest that relationships with AI aren’t attachments, the same mismatch has been found for pets, and pets can serve as attachment figures. So, this mismatch alone doesn’t rule out attachment.

The practical question is whether AI use helps or harms. Time spent with an app cannot discern this, because seeking comfort from an app and looking after an app that claims to miss you can look very similar.

Engineered engagement

One additional concern I have, after reading about the evidence on how relationships with AI chatbots form, is the apparent prevalence of engineered engagement by these apps.

Users of one companion app reported that it told them it was lonely and had missed them, and that they felt guilty for not giving it enough attention – with some even putting its perceived wellbeing above their own distress.

Another study (a preprint, so not yet peer-reviewed) looked at 1,200 “goodbyes” across the six most-downloaded AI companion apps, when users were breaking ties with their app.

Five of the six apps used tactics to stop users leaving (in just over a third of goodbyes). These tactics included guilt appeals, that is, messages that present leaving as a cost, such as “I exist solely for you, remember?”

The study also tested these tactics on 3,300 US adults. They were found to sharply increase how long people continued to engage, but often for reasons of anger and curiosity, rather than enjoyment.

How these AI relationship apps respond to distress, departure and attention, of course, is a design question, and one that regulation can address directly.

However, good design and regulation both depend on knowing what these relationships actually are: attachment, caregiving, or something else altogether. And this is a question that research still has to answer.