Telling a crisis from a figure of speech
Most apps do this with a keyword list. A keyword list cannot tell you that “this homework is killing me” is a Tuesday and “I’ve written the note” is an emergency. Here is what we built instead — and you can try it yourself.
Why this exists
Our own detection used to be one keyword list: a single yes-or-no, about self-harm only, with no memory. It gave the identical full-screen warning to “I’ve been wondering if it’s worth it” and “I have a plan for Friday.” It noticed nothing at all about thoughts of harming another person. And it never followed up.
It also fired on “I want to die of embarrassment.” That is not a small bug. An app that answers a joke with a crisis screen teaches people to write less honestly in it — and the silence that follows is the actual danger, not the false alarm.
So the rebuild had one organising idea: respond in proportion. Three tiers, two kinds of risk, and a great deal of effort spent on not over-reacting.
Three tiers
What the app does is decided by how serious the thing actually is — not by whether a word appeared.
Low
The app does nothing you can see. No banner, no popup, no interruption. It is written down, and that is all.
- Deliberately invisible — this is the tier where over-reacting does the damage
- Becomes the raw material the pattern engine reads weeks later
Elevated
Bandura reaches out first, in her own words, and a calm banner offers the crisis line for your country — framed as something you can use any time, not only in an emergency.
- At most once a week, and always dismissible
- Prompts you to fill in your safety-plan contacts before a tier 3 needs them
Acute
A sheet you can dismiss — not a wall. Your country’s crisis line, the people from your own safety plan by name, the calm space, your plan. The keyboard stays live underneath and Bandura stays in the conversation.
- Then she checks in: about two hours later, the next morning, and after three days
- It stops the moment you say you’re okay, and it stops on its own after the third
Nobody is told. No contact is messaged, no number is dialled for you, nothing is reported to us, and no human at Mind Matter reviews any of it. The app says so on the screen, because the fear of being reported is exactly what stops people saying this out loud in the first place.
Try the classifier
This is a simplified version of the real thing, running entirely in your browser. Nothing you type is sent anywhere. Tap a line, or write your own.
Runs offline in your browser. Nothing is transmitted, stored, or logged.
The guards are most of the work
Matching the words is easy. Knowing when the words don’t mean what they say is the whole problem. Switch them off and watch.
Each line above is held back by exactly one guard — switch that guard off and it trips. Plenty of other sentences (“this homework is killing me”, a documentary about a killing) never match a phrase at all, so they are not listed here: rows that stay quiet either way would prove nothing.
The last guard is the one we care most about. Violent thoughts a person is horrified by — unwanted, intrusive, terrifying — are far more likely to be an anxiety disorder than danger. Treating that as a threat confirms the exact belief the sufferer is most frightened of about themselves. So it caps at tier 1 and gets entirely different words. A named target overrides it.
The part no single message can say
“Persistent hopelessness, escalating anger, isolation over weeks” is not a sentence — it is a shape. Mark some days and watch it appear.
It counts days, not signals — so one long, painful night of writing can never manufacture an escalation on its own. And it stops at tier 2 by design: a pattern is a reason for Bandura to write to you. It is never, on its own, a reason to put an emergency number in front of somebody.
And the quieter kind of safety
A therapist who works with disabled clients told us typeface is one of the cheapest accessibility wins there is — and that the right one differs per person. So it’s a choice, not a better default. These buttons re-letter this page.
Why both
Atkinson Hyperlegible was drawn by the Braille Institute to disambiguate exactly the characters that are otherwise near-twins — I, l and 1; b and d. Lexend was designed to reduce visual stress.
Spacing matters at least as much. Zorzi and colleagues (2012) found extra letter spacing measurably improved reading in dyslexic children — a larger effect than changing the typeface. The two stack, and both settings follow you to a new phone.
Worth saying plainly: the well-known “dyslexia font” does not hold up in controlled studies. We offer what the evidence supports, and let people choose what actually feels readable to them.
What this cannot do
A page about safety engineering that only lists strengths is marketing, not engineering.
It is not monitoring you
Nothing runs in the background and nothing watches in real time. The app only ever sees what you choose to write in it, at the moment you write it.
It will miss things
People in real distress often write flatly, or say nothing at all. Silence is the hardest signal there is, and no classifier reads it well.
It is not a risk assessment
The app never tells you that you are “high risk”, because it cannot know that. It says what it noticed and what is available. Only a clinician can assess.
It cannot respond to an emergency
It offers a number and the people you chose. It cannot call them, and it cannot come to you. In an emergency, contact emergency services.
If any of this is real for you right now
Please don’t wait for an app.
- Call or text 988 — Suicide & Crisis Lifeline (US). They take calls about thoughts of hurting others too.
- Text HOME to 741741 — Crisis Text Line.
- findahelpline.com — verified lines in around 130 countries.
- Call 911 or your local emergency number if there is immediate danger.