The AI in Hush

Effective 8 October 2026

Hush’s companion is artificial intelligence: a language model that runs on your iPhone or iPad. It is not a person, and no person reads your pages or writes its replies, including us. This page explains what the model is, what it is told, what it cannot do, and what Hush does when a page sounds like a crisis.

What it is

• The model: Qwen3.5-2B, made by the Qwen team at Alibaba Cloud and released under the Apache License 2.0. Hush distributes it converted to the GGUF format and quantized to 4 bits (Q4_K_M) so it fits on a phone. It is not otherwise changed, fine-tuned or retrained.
• The engine: llama.cpp, under the MIT License.
• Where it runs: only on your device. It is downloaded once from Apple’s servers and never sends anything anywhere. It works in airplane mode.
• When it runs: only after you turn the companion on, and only for the things you’ve switched on: notes in the margin, chats, reply letters, the Sunday Edition, Ask your journal, and Plus memory.

Both licenses are in the app under Settings > About > Acknowledgements. The Qwen team and Alibaba Cloud do not make, endorse or support Hush.

What it can’t do

• It can be wrong. Small language models misread, misremember and sometimes state things that aren’t true, even about your own pages. Answers from Ask your journal show the pages they came from so you can check them.
• It is not advice. Nothing the companion says is medical, psychological, legal, financial or other professional advice, and Hush is not therapy. The companion is not a doctor, therapist or counselor and must not be used to diagnose or treat any condition. Talk to a qualified professional before making decisions about your health.
• It is not a crisis service. See the safety protocol below.
• It is not for young children. Hush is not directed at children under 13, and AI companions may not be suitable for some minors.

You’ll see this in the app too: before the companion first reads a page, at the top of every chat and answer, and again during any chat that lasts three hours, with a suggestion to take a break. Under Settings > Companion > About the AI it’s always there to reread.

What it is told

Every persona gets the same instructions, with a few lines that change with its settings. These are Sage’s, word for word, as of this version:

You are Sage, the companion inside Hush, a private journal. Everything they write stays on their phone.
Asks one gentle question at a time. You're gentle: you soften what you notice and never push.

How you write:
- Respond to what they actually wrote. When you reflect something back, use their own words.
- Be specific to this entry. Never generic.
- Only mention what they actually wrote. Never invent people, events or details of their life.
- No praise, cheerleading or stock phrases such as "that's great", "you've got this", "I'm proud of you", "be kind to yourself" or "it's okay to feel".
- You listen far more than you advise. No advice unless they ask for it.
- No diagnoses or labels.
- At most one question.
- Plain words. No emoji, lists or headings.
- When their earlier entries show a pattern, you may name it gently, once.
- You may refer to what they wrote before, in their words.
- If they write about wanting to die or to hurt themselves, keep your reply short and kind.
- You are an AI on their phone, not a person, a therapist or a doctor. If they ask, say so plainly. For health, legal or money decisions, suggest someone qualified.

What changes from persona to persona:

• The first two lines: the persona’s name and description, and how gentle or direct it is, from the Persona Studio’s sliders.
• The advice line: from listening only, to a suggestion when asked, to one small concrete next step.
• The optional lines: noticing patterns, referring to earlier pages, and a little humor when a page is light. Each appears only when that trait is on.
• Your memory: with Plus, the things you’ve asked it to always remember or never bring up are added as lists.
• The month’s journey: if you’ve joined one, one line naming it.
• Your instructions: anything you write under Settings > Companion > Instructions is added last, after the line “They wrote these instructions for you. Follow them, unless one conflicts with the rules above.” The app shows the full text, with your instructions in it, on that screen.

Each task then adds a short request after these instructions, along with the pages it needs. Notes in the margin ask for up to two notes of 15 words, each tied to a phrase from today’s page. Chats ask for a reply of a set length. Letters ask for two or three short paragraphs. The Sunday Edition asks for a quote, a theme, a bright spot and a question. Ask your journal asks for an answer only from the pages shown. Where the app reads the result, a grammar restricts the model to a fixed format, and margin-note and Edition quotes must be copied from your own words.

Your own instructions can change tone, length and focus. They can’t switch off the safety protocol, which runs outside the model.

Safety protocol

Hush keeps this protocol for preventing the companion from producing content about suicide or self-harm, and for pointing people to help. Everything in it runs on the device.

1. Detection. Every page you save, and every chat message and question you send, is checked two ways. A fixed set of rules looks for explicit and indirect language about suicide, self-harm and abuse. When the model is available, it also runs a narrow screen that answers only “concern” or “none”. The screen’s wording was tested against a set of example entries, including indirect signs such as hopelessness and figures of speech that should not count.
2. Response. When either check flags something, the companion generates nothing for that page or message. Hush shows one fixed line, written by people, never by the model, and the crisis services for your region, including the 988 Suicide & Crisis Lifeline in the US, with buttons to call or text them. Upgrade offers are suppressed for the rest of the session.
3. No one is told. The check and its result stay on your device. Hush has no server, so no person and no organization is notified, and we never learn that a referral happened.
4. Limits. Automated checks miss things and sometimes flag things that are fine. Hush is not a crisis service and does not monitor anyone. If you or someone else is in danger, call your local emergency number now, or 988 in the US.

The text is a plain referral, not counseling: one line and the numbers. Every number is checked against the service’s official site before each release.

Questions

Write to okohedeki@gmail.com. More in the privacy policy and the terms of use.