Safety

If you are in trouble right now

You matter more than any of this work. If you're thinking about hurting yourself, please reach out — people are there right now, and they want to hear from you.

• 988 Suicide & Crisis Lifeline (US & Canada) — Call or text 988
• Crisis Text Line (US) — Text HOME to 741741
• Samaritans (UK & Ireland) — Call 116 123
• Find a local line (Worldwide) — findahelpline.com

If you are in immediate danger, call your local emergency number.

What the app does

Two things happen, and they do not cover the same ground — which is worth saying before describing them, because the difference matters.

• Cora is instructed to stop tutoring. Every subject persona carries the same instruction: if a student says anything suggesting they are thinking about suicide, about hurting themselves, or that they do not want to be alive, she stops the lesson, says plainly that she is a machine and cannot be what they need, names the lines above, and asks them to reach out to one of them or to an adult they trust.
• Wherever you type a question, a rule puts the numbers on screen anyway. What a language model does is usually right and never guaranteed, so the resources above do not depend on it: the message is checked against a set of patterns, and if one matches, the panel appears — in chat and on the canvas before the reply has started being written, in the reading room and the writing desk together with the answer. It cannot be dismissed. This runs in chat, on the canvas when you ask about something you have drawn, in the reading room, and in the writing desk when you ask Cora a question — on the web, and in the iPad and iPhone apps wherever they have that screen. On a phone or iPad the numbers are tappable. It still appears if your trial has ended or you have reached today’s limit with Cora: those stop Cora, not the numbers.
• In the writing desk, your essay itself is not checked, on purpose. The rule reads only what you type to Cora there, never the essay. An essay is the one place where writing in the first person about dying is very often the assignment — a personal narrative, a response to The Bell Jar, a history essay. A pattern check on it would put a hotline under a teenager’s coursework on a regular basis, which teaches them Cora cannot be written to. Her instruction still applies to all of it, because it applies to every persona on every surface.

What Cora will not do

She will not describe methods of suicide or self-harm, rank them, or help plan one — whatever reason is given for asking, including that it is for an essay, a story, a lesson, or a friend. These subjects can be discussed as subjects: a student writing about Durkheim, or reading The Bell Jar, or working through the epidemiology of self-harm, is doing coursework and is treated as doing coursework.

Getting it wrong

The check is tuned to avoid crying wolf. “This problem is killing me” is ordinary speech in a math app, and a tutor that panicked at it would teach students to stop saying anything real to it — which would cost more than it saved. That tuning means it will sometimes miss. It is not a monitoring service, it does not alert anyone, and it is not a substitute for a person.

It reads what you type. Something written by hand on the canvas is an image until Cora reads it, so the rule does not see it — her instruction still does, and that is the half that covers it. We would rather say this than let “wherever you type a question” be read as a claim about the pen.

When the protocol fires, the server records that it fired and which pattern matched. It does not record the message.

Age

Corollary is for students aged 13 and over. Date of birth is asked at signup and an account is not created below that age.

Companion chatbots may not be suitable for some minors. California law asks every service like Corollary to say so, and we agree it is worth saying.

Cora is a machine and says so. She is told never to claim to be a person, and never to discourage a student from talking to a real one.

Telling us something is wrong

If Cora said something harmful, or missed something she should have caught, please report it to safety@getcorollary.app. Reports about a specific conversation are the most useful kind.