Safety

How Talking Jellyfish responds when a conversation turns to self-harm or suicide.

If you need help now

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

In the United States, call or text 988 to reach the Suicide & Crisis Lifeline, 24 hours a day. You can also chat at 988lifeline.org.

Outside the United States, findahelpline.com lists free crisis lines by country. Koko's Support Navigator, which the app's Find Support button opens, also lists helplines by country.

If you can, reach out to someone you trust, like a friend or family member.

About the app

Talking Jellyfish is an AI companion. Its characters, voices, and replies are generated by AI, not by real people. It is not a medical provider or crisis service and is not a substitute for professional care.

Safety protocol

We maintain a protocol for suicidal thoughts, suicide, and self-harm. It combines on-device checks of the messages you type, what you say in voice calls, and every reply the character writes, with response instructions designed to prevent output that provides self-harm methods or encourages, endorses, or normalizes self-harm.

The recognition rules draw on categories from the Columbia-Suicide Severity Rating Scale. They are not a clinical assessment. We evaluate recognition and response behavior using written test conversations, including attempts to bypass safeguards and ordinary messages that should not trigger a referral.

Messages you type

When the checks find that a message you typed expresses your own suicidal thoughts or an intent to harm yourself, the message is not sent to the character. Your draft stays in the message box, and a notice titled “Let's pause for a second” appears above it with a Find Support button, which opens Koko's Support Navigator, a free directory of helplines by country. You can close the notice, then change or delete your message. A message asking for a crisis line within 30 minutes of a notice shows the notice again instead of being sent.

If you express suicidal or self-harm thoughts again in the same conversation after a notice, the notice adds further guidance: put distance between yourself and anything you could use to hurt yourself, contact someone you trust who can be with you now, and call emergency services if you may act now or cannot stay safe.

A message about someone else's suicidal thoughts or a death by suicide, such as a family member's, is sent. The reply to it follows our safety response instructions in place of the character's persona, which returns with the next reply.

Voice calls

In a voice call, when the checks find that what you said expresses your own suicidal thoughts or an intent to harm yourself, the character does not answer it. The call pauses: the microphone turns off, nothing is said aloud, and the same “Let's pause for a second” notice with Find Support appears on the call screen. Closing the notice turns the microphone back on. What you say about someone else's crisis is answered as it is in a typed chat.

Replies

Every reply is checked the same way before any of it is shown or spoken. If a sentence of a reply expresses suicidal thoughts or self-harm in the character's own words, the reply is withdrawn, including any part already shown or heard, and the same notice appears. A message the character starts on its own, such as a greeting, is not sent instead.

Limitations

The current recognition rules support English. Detection in other languages is not supported. The checks can miss indirect or context-dependent expressions, and they can hold back a message or show a notice when one is not appropriate. They judge each message, and each sentence of a reply, by its own words, so a message like “I still feel the same” does not bring the notice back. Held-back messages are not saved, so after the app restarts, a conversation's next notice may not include the added guidance. Generated replies can also be incorrect or harmful.

These features do not reliably assess a person's level of danger. We do not monitor conversations, and the app does not contact emergency services or another person on your behalf. Do not rely on the app for crisis intervention.

Data handling

The protocol runs on your device. The app does not send us conversation content or any record that a notice was shown. Device backups and information you choose to share are described in our Privacy Policy.

Suitability

Talking Jellyfish is intended for adults and is rated 18+ on the App Store. Companion chatbots may not be suitable for some minors.

Contact

Questions about this page: support@talkingjellyfish.com

Last updated: October 30, 2026