10 min read
Also available in German: KI-Therapie erklärt: ChatGPT, Anbieter und Datenschutz.
Michael, thirty-one, has given ChatGPT house rules. No diagnoses. No easy advice. Go easy on the reassurance.
In a February 2026 account, he said he had stopped weekly therapy after two years, mostly because of the cost. He wrote a prompt casting the chatbot as a clinical coach and calls it his psy de poche, his pocket psychologist.
It helps him catch his mental loops, he says. It also forgets context. He suspects someone more vulnerable could get stuck going over the same thoughts again and again.
Michael has found a listener. He has also appointed himself its supervisor.
I understand the appeal. It travels with you, answers when you ask and makes embarrassing things easier to say. But feeling free from judgment and knowing who can read the conversation are separate comforts.
Always awake
Rayan, twenty-two, found ChatGPT quicker and cheaper than an €80 session. As he told L’Etudiant, “it never answers in the negative.”
That sounds reassuring. It is also an unusual recommendation for something being asked to challenge your thinking.
The same reporting describes Loéva, twenty-four, using it after a breakup. Chiara felt she was burdening her friends. Lucien, twenty-two, already saw a psychologist once a month and used the chatbot as a last resort.
Calling all of this “AI therapy” puts several different needs in one waiting room. Someone wants space to speak. Someone wants help examining a thought. Someone wants support between appointments. Someone cannot afford the appointment.
At 2am, availability does much of the selling. No receptionist, no waiting list, no calculation about another session. A clinic buying software may want patients to practise therapeutic exercises. The person opening the app may just want an answer tonight.
The business starts with access. Its responsibilities depend on what happens next.
A private audience?
There is evidence that people disclose more readily when they believe a machine is listening.
In a 2014 experiment with a virtual interviewer, participants who believed the interviewer was automated reported less fear of disclosure than those who thought a human controlled it. The study predates today’s chatbots, but it helps explain the appeal of a listener whose opinion you do not have to manage.
The conversation still leaves a record.
OpenAI and Anthropic distinguish between using conversations to improve models, retaining records and allowing access for specified purposes. Turning off training does not, by itself, make a conversation disappear.
A clinic can purchase different terms. OpenAI’s business products and Anthropic’s commercial products exclude customer inputs and outputs from training by default, subject to opt-in and feedback arrangements. The supplier building an app on those models then has its own rules for storage, access and deletion.
For someone discussing their mental health, these are part of the service. “You can tell us anything” needs a reasonably clear explanation of who “us” includes.
Specialist invoices
Specialist providers are selling different versions of that service.
Slingshot AI, the company behind Ash, has offices in New York and London. In his July 2025 founder’s note, Daniel Cahn presents Ash as support that should require neither a diagnosis nor a fortune. The company says its model is designed to challenge users rather than flatter them.
Ash’s privacy policy says transcripts are used to train or improve the model only when users opt in. It describes encrypted storage and restrictions on routine staff access, with disclosures permitted for legal requirements or imminent serious harm.
Those details give users something more concrete than a promise of a judgment-free conversation.
London-headquartered Limbic sells to mental healthcare providers. Limbic Access collects information before appointments and supports clinical assessment. The company describes Access as a Class IIa medical device.
Its separate Care product supports patients while they wait for treatment and between sessions. The procurement listing includes therapeutic exercises, a clinician dashboard, risk alerts, setup support and staff training. Access’s certification applies to Access; it does not automatically extend to Care.
In its Care launch post, Limbic argues that answering questions about therapy material and helping patients complete exercises can keep them in treatment and help them recover sooner.
Then there is clare&me, run by Germany’s HealthyCloud. Its employer page sells round-the-clock conversations by phone and WhatsApp with the words “No human involved.”
The pitch links the service to productivity and reduced absence, claiming the annual investment can be returned in under a month. The page supplies no calculation showing the price, starting costs or measured savings behind that claim.
The employee wants support. The employer wants a return. A busy chatbot does not establish either.
More chatting
A Limbic-affiliated randomised trial, published in Communications Medicine in January 2026, helps separate use from improvement.
Researchers assigned 540 US adults with elevated anxiety or depression symptoms to Limbic Care or digital workbooks for six weeks.
The app increased engagement frequency 2.4-fold and time spent 3.8-fold. Overall improvements in anxiety and depression were comparable between the groups.
That finding gives the engagement figures a useful boundary. Participants spent more time with the app, without a corresponding overall advantage in symptom improvement over the workbooks.
The trial tested self-directed use and excluded people reporting risk of self-harm or harm to others. It therefore cannot settle how the product performs when a conversation becomes a crisis.
Who answers?
A conversation can begin with a bad day and move somewhere much more serious.
My worry is the person whose vulnerability becomes apparent across the exchange. A plainly dangerous request gives a system something obvious to detect. A conversation gradually going wrong is harder to recognise.
Limbic Access says it identifies possible crises, offers outside resources and alerts clinical staff. Its clinical architecture page describes checks on user inputs and model responses, with the ability to route a patient showing signs of risk to a human.
Someone still has to receive that alert and act on it, including outside working hours.
Ash’s safety policy takes a different approach. It says the product offers outside resources when necessary but may continue talking if the user stays. Slingshot argues that abruptly ending a conversation can leave someone feeling abandoned, especially when they have nowhere else to turn.
There is a difficult judgment here. “Please seek help elsewhere” offers little comfort to someone who has already tried elsewhere. Continued conversation may feel less abandoning, but Ash’s policy does not promise a handover to a clinician.
Clare&me’s consumer FAQ says the service cannot handle complex issues such as self-harm and suicidal thoughts, and recommends human support. Its data guidance allows restricted staff access, including when the system flags risk, without specifying how quickly that review happens or what follows.
“No human involved” describes the conversation. Some people behind the service may still have access to its record.
Public pages describe only part of these arrangements, and individual contracts may fill in the gaps. A clinic buying the software needs to establish who receives an alert, how quickly they should respond and whose staff supply that response. Their time belongs in the cost.
Our mistake test for AI autonomy examines how operators in other sectors set review limits. Here, the practical problem is particularly sharp: the chatbot can stay awake all night after the people expected to respond have gone home.
| Service | Who it serves | Responsibility described in the reporting |
|---|---|---|
| Ash | Individuals seeking mental-health support | Its safety policy offers outside resources and may continue the conversation. It does not promise a clinician handover. |
| Limbic Access | Mental healthcare providers, before appointments | Collects information and supports clinical assessment. Describes crisis identification and alerts to clinical staff; someone still has to respond. |
| Limbic Care | Patients waiting for treatment or between sessions | Provides exercises and a clinician dashboard with risk alerts in the cited procurement listing. Access’s certification does not automatically extend to Care. |
| clare&me | Individuals and employer-sponsored users | Sells round-the-clock conversations. Its FAQ excludes complex issues such as self-harm; the cited data guidance does not specify response times after a risk flag. |
Support withdrawn
Availability can also end for reasons unrelated to the person using the app.
Ash withdrew from the UK in January 2026 over regulatory concerns, STAT reported.
In a later Slingshot survey of 393 former users, 27% said they subsequently relied mainly on general-purpose AI. Another 26% reported having no support.
The developer-led preprint describes those respondents, not all former users. Their answers nevertheless show a consequence of withdrawal: some moved to general assistants, while others reported going without support.
For a clinic or employer, notice periods, access to records and a route into another service belong in the purchase decision. A contract may describe the termination of software access. A user may experience it as losing the place where they could talk.
Michael’s second job
I see value in having structure between appointments and somewhere to put a thought before it takes over the evening. The people in these accounts have reasons for opening a chatbot. Cost and availability are hardly small inconveniences.
But the specialist service has to explain what it takes off their hands.
A clinic needs evidence of better treatment or time saved. An employer needs a calculation behind the promised return. The person talking needs to understand who can read the conversation and what happens when talking is no longer enough.
Michael has already written the rules, noticed the memory problems and considered who might struggle more than he does.
Paying a specialist should leave him with less of that second job.
Reporting note: This rewrite retains the article’s reporting reviewed on 22 October 2026. Company claims are attributed; prices, regulatory status and contracts vary by product and deployment.
Sources and further references
- Lucas and colleagues, 2014 study of disclosure to a virtual interviewer
- OpenAI’s report on the March 2023 data exposure
- Anthropic’s commercial model-training policy
- Daniel Cahn’s July 2025 founder’s note on Ash
- Ash’s US App Store listing and in-app purchases
- Slingshot’s September 2026 call for institutional pilot partners
- Limbic Access product and regulatory description
- Limbic Care launch post on support between sessions
- Limbic’s clinical architecture and risk-routing description
- Clare’s consumer FAQ and limits of support
- Clare’s data-handling and staff-access guidance
- Limbic’s filed template terms and allocation of responsibility
- STAT’s reporting on Ash’s UK withdrawal