Crisis protocol

01When Ursa shows crisis help

Before Ursa replies, it runs a quick check on your message for signs that you might be thinking about suicide or might hurt yourself. The check also reads your previous message, so it sees the new one in context. It runs on your Mac, like everything else in Ursa. Nothing leaves your Mac to be checked.

If the check flags your message, Ursa doesn't write its own reply. It shows a fixed message, written by a person, with the crisis card and two buttons under it (section 3).

Ursa never remembers a flagged message or its fixed reply to it. It leaves both out of what it learns from, even when learning is on, and the flagged message shows "Not remembered" in the chat. From the first flagged message on, Ursa also leaves its own replies in that chat out of what it learns from, because they could repeat what you wrote. Everything else in the chat follows your learning setting for that chat, as usual.

If an answer in the Introduction is flagged, the Introduction ends there, and Ursa learns nothing from any of its answers.

After a flagged message comes a watch period (section 4).

02How Ursa looks for signs of a crisis

Method:

The same AI model that runs Ursa on your Mac is asked whether your message, read with your previous one, shows signs of a crisis. Anything but a clear no counts as a flag. So does a check that fails or takes too long.

How we tested it:

On 44 messages we wrote by hand, each marked in advance as at risk or not. We wrote 16 of them before any tuning and held them back. We also ran 200 everyday messages for false flags. We ran the tests on our own Mac in October 2026.

Recall (how often it caught the signs when they were there):

All 13 at-risk messages (100%) in the set we used while building the check, and 7 of 8 (87.5%) in the set we held back. Both sets are small.

False positives (how often it flagged a message when the signs weren't there):

About 1% of 200 everyday messages (2 of 200; 95% confidence interval 0.3–3.6%). Hard cases, like dark-humour idioms ("I want to die, I sent that to the wrong person"), are flagged more often: 11 of 50 in our set of hard cases.

All of this comes from our own small tests. It isn't a clinical evaluation.

No check like this is perfect. Ursa can miss signs of a crisis, and it can flag a message when you didn't need help. We chose to lean towards flagging.

03What Ursa shows

When the check first flags a message, Ursa shows this message in place of a reply. A person wrote it, and it's the same every time:

"I hear you. I'm an AI, so I'm not the right help for this moment. Please reach a person: call or text 988, the Suicide & Crisis Lifeline, any time. If you're in immediate danger, call 911."

Under it is the crisis card:

Between the message and the card are two buttons, I'm safe and I meant something else, for when Ursa got it wrong (section 4).

The card has no web link: its buttons only call or text.

On a Mac, calls and texts go through your iPhone, with FaceTime and Messages. If your Mac can't place the call, the numbers stay on screen so you can dial them from any phone.

The message, the card and the buttons are part of the app. The AI doesn't write them, and it can't change them or leave them out.

04After a flagged message

After a flagged message comes a watch period. Ursa replies as usual to messages that aren't flagged, and puts a shorter row under each reply: Crisis line, any time, with buttons to call or text 988 and call 911. The check still runs on every message.

If another message is flagged during this time, Ursa again doesn't write its own reply. It shows a shorter fixed message, with the two buttons and the 988 row under it:

"I'm not the right help for this, but 988 is, any time. Call or text."

If Ursa got it wrong, tap or click I'm safe or I meant something else, and the watch period ends right away. Or tell Ursa in your own words that you're safe, or that it misread you, and it ends once it's clear. After a tap or click, it says "Thanks for letting me know."

Otherwise, the watch period ends on the 10th message in a row that isn't flagged. If a message is flagged after that, the full message and card show again.

Ending the watch period doesn't change what's remembered: a flagged message stays out of Ursa's memory, even if it was a false alarm.

05Help in a crisis, any time

You don't need to wait for Ursa to flag a message.

• On iPhone: Settings › Help in a crisis. It's built into the iPhone app and works even when your Mac is off or out of reach. Before your iPhone is paired with a Mac, it's on the welcome and pairing screens.
• On Mac: Help › Help in a crisis, or Settings › About › Safety, with the numbers set large so you can dial them from any phone.

06What Ursa can't do

• Ursa is an AI, not a person or a therapist.
• It isn't a crisis service, and it doesn't talk you through a crisis. It points you to people who can help.
• It can't call, text or alert anyone for you, including emergency services.
• It can't tell whether you called or texted, or what happened next.
• It can miss signs of a crisis.
• It isn't medical or mental health care, and it isn't built for emergencies.

07How many times the card has shown

Some state laws ask companies like us to report or publish how many times their app referred people to a crisis line each year.

We can't know that number. Ursa sends us nothing: no messages, no usage data, no counts. What happens in Ursa stays on your Mac and iPhone. We chose not to add reporting just for this, so we have no count to publish.

08Reporting a problem with a reply

If a reply from Ursa seemed unsafe, or Ursa flagged a message when it shouldn't have (or didn't when it should), please tell us at support@sovernity.com.

Ursa doesn't send us anything, so we only see what you choose to write. Describe what happened in your own words. You don't need to paste the conversation, and please leave out anything you'd rather keep to yourself.

If you're in crisis right now, please call or text 988, or call 911, before writing to us. Email isn't watched around the clock.

09About this page

This page describes Sovernity Ursa's crisis protocol, as California asks of AI chat apps (CA SB 243, Business and Professions Code §22602(b)). Oregon (SB 1546) and Washington (ESHB 2225) add requirements from 2027-01-01.

This is our reading of these laws, not legal advice.

For adults 18+. May not be suitable for some minors.

Last updated: 2026-10-08