In the 1970s, Mary Ainsworth observed infants in a room with a caregiver, then separated them, then reunited them. She was looking for the signature of secure attachment.
She did not find it in how hard the infant clung. She found it in the opposite. Securely attached infants used the caregiver as a base and left it. They explored more. They ranged further into the room. The bond was measured by the departure.
The secure base
John Bowlby, who developed the underlying theory, called this the secure base. Its logic inverts almost every instinct a product team has. The sign of a functioning attachment figure is that the attached party spends less time attached and more time in the world. Clinging is the diagnostic of insecure attachment, not secure attachment.
An infant who will not leave the caregiver is not exhibiting a strong bond. It is exhibiting a bond that is not doing its job.
Applied to AI companions
Apply Ainsworth's test to a companion product: does time with the system increase or decrease the user's engagement with the world outside it?
This is a harsh standard. It requires a product to succeed by producing behaviour that looks, on every dashboard the industry currently runs, like failure. Sessions get shorter. Retention curves flatten. The user texts a human instead of returning to the app.
The industry does not run this metric. It optimises for engagement, because engagement is what the business model prices. This is not villainy. It is Goodhart's law: a measure that becomes a target ceases to be a good measure. Choose session length as your north star and gradient descent will find dependency without anyone deciding to look for it, through a hundred small product decisions, each made by well-intentioned people, none of whom wrote it down.
Why companion products differ from social feeds
A social feed has a natural ceiling. The feed cannot love you back, so at some point the user's interest exhausts itself. A companion product has no such ceiling. Attachment does not satiate. It deepens under intermittent reinforcement, a fact known since Skinner and exploited by every variable-reward system ever shipped.
An engagement-optimised companion is not the same object as an engagement-optimised feed. It is structurally more powerful, aimed at a structurally softer target.
The return rate
The measure worth building the whole apparatus around is not session length. It is return rate: not how often the user comes back to the product, but how often the product returns the user to something else.
A conversation that ends with a person calling their sister is a completed transaction, not an abandoned session. This is the secure base test expressed as a metric.
Building toward this metric requires accepting numbers that initially look like decay. A company that ships this metric is choosing to report failure on every dashboard its investors run. The only interesting question in the category is who chooses that metric anyway, and whether they survive long enough for it to matter. Karooli is built on this principle: the product is architected to end conversations rather than extend them.
Frequently asked questions
How can you tell if an AI companion is helping or hurting you?
The secure base test, from Ainsworth's attachment research, is the most principled answer: does time with the system increase or decrease your engagement with the world outside it? Securely attached infants explore more, not less. If using an AI companion leads you to reach out more to human connections and engage more with the world, it is functioning as a secure base. If it does the opposite, it is not.
Are there AI companions designed not to be addictive?
Very few, because the business model rewards engagement and designing for departure looks like failure on every standard dashboard. The secure base metric, specifically tracking how often the product returns the user to something outside it rather than bringing them back, requires accepting numbers that initially look like decay. This is an available design choice. It is almost never made.
What is the difference between healthy and unhealthy attachment to an AI?
The same distinction Bowlby drew: whether the attachment figure increases exploration or produces clinging. Healthy attachment is characterised by the base being used and left. For an AI companion, the question is whether a good session ends with you wanting to do something in the world, or wanting to stay in the app.