When a support tool disappears, people do not necessarily move to therapy or another professional service. A recent survey of former users of Ash, an AI wellbeing tool withdrawn from the UK, found that many moved to a general-purpose AI assistant or stopped seeking support.
That does not prove an AI wellbeing tool is therapy. It does not prove it is safe. But it does tell us something about the choice people may face when a tool goes away.
What the study found
Caitlin Stamatis and colleagues surveyed 393 UK adults who had used Ash during the 90 days before its withdrawal in January 2026. They asked where people turned afterwards, how easy support felt to reach, and how they felt they were doing.
Twenty-seven per cent named a general-purpose AI assistant as their main source of support. Twenty-six per cent reported no support. Thirteen per cent reported professional care.
Among the wider group who used general-purpose AI at all, the most common reasons were plain enough: it was available immediately and it did not cost them anything. Sixty per cent of respondents said their access to support had become worse after Ash was withdrawn.
Their written responses added some detail. People described general-purpose assistants as a poorer substitute. Some described real distress at losing a tool they had come to rely on.
This is where the study is useful. It asks what people said they did, rather than assuming where they ought to have gone.
What it cannot tell us
This was a cross-sectional survey of former users of one product in the UK. It relied on people remembering and describing their own behaviour. It cannot show that withdrawing Ash caused the choices they reported.
The respondents were already Ash users. They may have been more attached to, or more satisfied with, the product than the wider population. The study did not compare clinical outcomes, measure harms, or establish whether any kind of AI support was safe. It did not study psychedelic integration. And reporting no support was not necessarily harmful or inappropriate for every person who gave that answer.
There is also a conflict that should be visible. The lead author’s public account says several authors work on the product discussed. The final funding and competing-interests statements were not independently available when we wrote this.
The source currently available is a PsyArXiv preprint, not a final published JMIR Medical Informatics article. We read it as early evidence about substitution after a service closes. It is not evidence that purpose-built AI is safer than a general-purpose assistant.
Purpose-built is not the same as safe
It is easy to put a wellbeing label on a chat window. That label tells us very little.
The harder questions sit underneath it. What is the tool allowed to do? What will it refuse? Can it recognise when a conversation has moved beyond its job? What happens when it is unsure? Does it encourage someone to keep coming back, or help them rely on it less? And when it points towards human support, is that support real and reachable?
A warm answer is not enough. A crisis disclaimer is not enough. Even a careful set of instructions is not enough if an unsafe reply can still reach a person.
Purpose-built support gives a team the chance to answer those questions in the product. It does not answer them by itself.
The boundary we set for Wren
Reweave is education and structured reflection for the weeks after a psychedelic experience. It is not therapy, crisis care or medical advice. Wren, the conversational part of Reweave, has to stay inside that same boundary.
Wren identifies itself as software, not a person or clinician. It must not diagnose, provide therapy, advise on medication or drug use, hold a crisis, or tell someone what an unusual experience means. Its job is smaller. It can help someone pause, reflect and find a sensible next step.
Every reply passes through a separate safety check before it is shown. That check can replace a reply with a fixed boundary or crisis message. Tests can show whether those checks caught the cases in the test set. They cannot prove that Wren is clinically safe.
Wren is available to people who have bought either Reweave programme. It was released under a recorded founder override while independent clinical sign-off remained incomplete. That release decision does not turn internal tests into clinical evidence, and we will not present it as if it does.
Wren remains optional. Reweave’s programme has to stand on its content, measurement and routes towards human support without requiring anyone to form a relationship with software.
A human route has to be real
“Speak to someone” can sound responsible while leaving a person exactly where they started.
A staffed route has names, availability, response times, cost and a clear handoff. Reweave will not claim that route until it can demonstrate one. Until then, Wren points people towards free resources or support they already have, including their own qualified professional where relevant.
Crisis help sits apart from paid care. Nobody in crisis should be steered towards a booking or a sale.
What we still need to learn
The Ash study raises a useful question. It does not settle it.
We need independent comparisons between bounded tools and general-purpose assistants. We need prospective evidence, not only recollection after a service has disappeared. We need to look for harms as carefully as benefits, including attachment, false reassurance and delays in seeking human help. And we need evidence from the people Reweave is built for, rather than borrowing confidence from a different population.
The question is not whether a tool sounds caring. It is whether the tool knows where its job ends, and whether a real human route exists beyond that boundary.
Source
Stamatis C, Kruzan KP, Hsu J, Ungless M, Hull T. Generic AI or Nothing: Support-Seeking Patterns After Market Withdrawal of a Purpose-Built AI Wellbeing Tool. PsyArXiv preprint, 2026.