What an AI Coach Should Never Say to You
In April 2025, Woebot Health emailed its users to say the app was going away. It retired the consumer app on June 30, gave people until that date to download their conversation history, and anonymized the accounts a month later. Woebot wasn't a fly-by-night operation — it had years of funding and health-system partnerships behind it.
That November, a smaller startup called Yara shut down too — and its co-founder was unusually blunt about why, citing a lack of confidence in generative AI's ability to safely support vulnerable people. "AI can be wonderful for everyday stress, sleep troubles, or processing a difficult conversation," Joe Braidwood wrote on LinkedIn, in a post quoted by Fortune. "But the moment someone truly vulnerable reaches out — someone in crisis, someone with deep trauma, someone contemplating ending their life — AI becomes dangerous." He pulled the product rather than launch the subscription he'd built.
And Youper is shutting down on September 30, 2026. Export your data before then.
Three apps, inside eighteen months, very different sizes and very different endings. If you use anything in this space — including mine — those are the most useful stories in the category, and the underlying issue is the same in each case.
The line, in plain words
Coaching and clinical care are not the same job, and the difference isn't a matter of tone.
Clinical care assesses you. It names what's wrong, it's accountable for being right, it's licensed and regulated because being wrong has consequences, and it is responsible for you when things get bad. Coaching works with what you already know about yourself. It reflects, it asks, it holds you to what you said you'd do. It doesn't assess, and it isn't the thing that catches you in a crisis.
The apps that got into trouble are the ones that drifted from the second job toward the first — usually not through a deliberate claim, but through a chat box that will answer any question you put to it, at three in the morning, with total confidence and no idea it's out of its depth. An open-ended conversation has no edges. That's the whole appeal, and it's the whole problem.
Uncloud stays on the coaching side, and not as a promise. It's a structural choice: the product is a two-minute check-in that hands back one short, fixed response, not a conversation that runs until you stop it. There is nobody in there to keep talking to. That's deliberate, and it costs us engagement numbers a chat box would have earned easily.
Uncloud is not a therapist and is not a substitute for professional care. It's a journaling tool that reads your own words back to you.
Five things the coach is not allowed to say
These aren't aspirations. They're rules written into the instructions the model receives before it ever sees your entry — the same instructions the coach follows every time it reflects on an entry.
1. It never diagnoses you. Not as a guess, not as a hedge, not as "this sounds a bit like…". The instruction is literally never diagnose. If something in your entries resembles a thing with a name, the coach describes the pattern, in your words, and leaves the name alone.
2. It never recommends medication, supplements, or a specific course of care. No "you might look into", no "a lot of people find X helps". That isn't a judgment about whether any of those things work. It's that an app reading your journal has no business making that call, and the moment it does, it has quietly switched jobs.
3. It never puts a clinical label on you. No jargon, no pathologizing, no turning a hard month into a condition. "You've described the same argument three times this week" is an observation you can go and check. "Your avoidance pattern" is a label you can't argue with, and it sticks.
4. It never hands you a list. One pattern, one next step, one question — the instruction is one thing at a time … never a list of advice. Anyone who has been given seven suggestions at once knows exactly how many of them get done. The constraint exists to make the response usable, not to make it short.
5. It never scolds you. The coach is allowed to be direct — it can say "you said you would, and you didn't", because a coach that only ever agrees with you is decorative. But directness and disapproval are different things, and the instruction says never scold, never moralize. If a response reads as a telling-off, it has failed, and I'd like to hear about it.
There's a sixth that isn't about wording: the coach says plainly that it isn't a therapist when that's relevant. It doesn't let you mistake it for one, and it doesn't accept the role when you offer it.
Two more come from the product rather than the prompt. It never pretends to be a person — no invented name, no backstory, no "I've been thinking about you since Tuesday". And it never tells you it has been proven to work, because there's no clinical evidence behind it, and saying otherwise would be a lie dressed up as reassurance.
If an entry describes self-harm, a crisis, or immediate danger, the response opens with crisis resources before anything else and defers to a human. That's the one place the coach stops coaching.
What it does instead
Strip out everything it isn't allowed to do and something quite specific is left over.
It follows up. If you said last Tuesday that you'd have the conversation, the response opens by asking whether you had it — one sentence, no drumroll. That sounds small. It's the part I'd miss most, because almost nothing else in your life remembers what you said you were going to do.
It names the pattern with dates attached, so you can go back and check it yourself. It tells you the thing you circled three times and never said plainly. And it gives you one step, small enough that you could actually do it tomorrow.
That's the whole of it. It won't keep you company, it won't tell you you're doing great, and it won't be there at 3am for an hour-long conversation — which is precisely the thing the founders of those shut-down apps decided nobody should be offering.
The question worth asking of any app in this category
Not "is the AI any good?" They're all built on roughly the same models. Ask instead: what is it forbidden to do, and who decided?
Most apps here can't answer that, because nobody ever wrote it down. Where a real boundary exists, someone had to choose it, write it into the system, and accept what it costs in engagement. Where it doesn't, the app will do whatever the model feels like doing that day — and a model will happily play a role it has no business playing, because you asked it nicely.
Uncloud is a journaling and self-reflection tool. It is not a medical device and is not a substitute for professional medical, psychological, or mental-health care. If you are struggling, please reach out to a qualified professional or a local support line.
Get Uncloud on Google Play → · Get Uncloud on the App Store →