AI safety — Zenith Guard
An assistant that can open apps and speak answers needs a guard at the door. Zenith Guard is that guard — a check that reads every request and every reply from ask the sky — the launcher's built-in AI assistant — against the suite's content policy.
It checks both ways, and it runs on your device. Before your question reaches the AI, Zenith Guard reads it; a blocked request never reaches the model at all. Before a reply is shown or spoken, it reads that too; a flagged reply is replaced by a short refusal. Nothing about this check leaves your device.
A refusal names the policy it fell under: "I can't generate a response to this request. (Zenith safety policy: <category>.)" A refusal about self-harm instead points you to the 988 Suicide & Crisis Lifeline and to someone you trust.
Why a reply carries a caveat
Below the conversation sits a small note: "✦ AI-generated — verify anything that matters." An assistant can be confident and wrong, so the note is a standing reminder to check things that count. Beside it sit links to "acceptable use" and "ai & safety".
Reporting, and how it teaches the guard
Each reply has a "⚑" report action. Tap it and you see "reported — thank you". Reporting does more than flag — it quietly tightens your own guard. A few phrases from the reported reply are added to a list on your device, so a similar reply is blocked next time. This only ever blocks replies — never what you ask, and never the words of the ask that reply answered. It stays on your device; nothing is uploaded.
Every block and every report is also written to a safety log on your phone. Only the newest 14 days are kept.
What it looks for
The checks weigh content against the full suite policy — things like harassment, hate, dangerous acts, deception, and self-harm, along with attempts to jailbreak the assistant, plus categories added for app-store parity such as deceptive election content, malicious code, and fake official documents.
Where to see it
Open Settings › ai safety. The page shows two on-device counts — "blocked requests" and "your reports" — with "copy safety log" (it copies the log to your clipboard), "refresh", and links: "acceptable use policy ↗", "ai & safety ↗", and "privacy policy ↗". "blocked requests" counts every block — a request stopped before the AI, and a reply stopped before you saw it. Both counts cover the 14 days the log keeps. Everything here is read from your device; nothing here leaves it.
One last distinction: Zenith Guard checks content. Separately, the first time you send to your AI provider, a one-time disclosure asks your permission to contact them at all — a different gate, covered with setup. You can withdraw that yes later in Settings › privacy & permissions, under "third-party AI consent".
See also. Ask the Sky — overview · What Ask the Sky can do · AI safety — Zenith Guard & consent · Privacy & permissions · Set up Ask the Sky