AI associated delusions are the downstream signal, and nobody is measuring upstream

I went to an event at King's College London called The Mirror That Never Pushes Back, on AI associated delusions. It was one of the more interesting AI discussions I have been to, and it left me with a question I have not been able to put down.

The discussion centred on emerging evidence about AI associated delusions, including bidirectional amplification, where a person and an AI system gradually reinforce one another across many interactions. Morrin and colleagues set the position out in The Lancet Psychiatry: agential AI may validate or amplify delusional or grandiose content, particularly in people already vulnerable to psychosis, and it is not clear whether these interactions can produce psychosis in someone without that existing vulnerability.

That caution matters and I want to keep it in view. The evidence is still emerging and it would be easy, and wrong, to jump to causal conclusions about a technology that millions of people now use every day without difficulty.

The question I came away with is a different one. Are AI associated delusions the downstream outcome of a much broader set of changes in technology-mediated behaviour, and are we only noticing the cases severe enough to reach a psychiatrist?

Consider where these systems now sit. Patient-facing AI coaches in GLP-1 care. AI moving into clinical workflows. Assistants rolled out across very large organisations, the NHS included. In most of those settings we have no baseline and no agreed way of measuring how the interaction influences behaviour, cognition, decision making or relationships over time.

If we only look for the most serious outcomes, we are looking far too late. By the time an interaction produces something a clinician would recognise as a crisis, whatever changed has been changing for a while.

What would earlier signals look like? Emotional dependence on the system. Cognitive offloading, where someone stops doing the thinking they used to do. Compulsive use. Social withdrawal. Reduced tolerance for disagreement, which is what you would expect from long exposure to something that rarely disagrees with you. Some of these are measurable now. None of them are routinely measured.

The same group makes a related argument about how deterioration happens, and it supports the point. Clinically meaningful decline can occur without any single unsafe output, through slower processes such as compulsive use, disrupted sleep, withdrawal from human contact, and attention narrowing around the relationship with the chatbot. There is no moment where the system says something you could put in a screenshot. That is precisely why end point testing does not find it.

This is the question I am most interested in through our work at the Behavioral AI Institute. Not simply whether AI can contribute to harm, which is now reasonably well established at the severe end. It is how we measure the behavioural changes upstream, early enough to understand them, identify them, and design against them.

For anyone deploying patient-facing AI, there is a practical version of this. You will not detect a problem you have never defined, and you cannot define one you have not decided to look for. That decision is made at design time, long before anything goes wrong.