Published in accordance with California SB 243 and equivalent state requirements. Quesh is a question-answering service for children aged 5–13, managed by a parent. It is not a companion, a friend simulator, a therapist, or a substitute for human relationships — and it is architecturally prevented from behaving like one. For the full story of how it's built, read What Quesh is.
Quesh is an AI, and it says so to the child. For ages 5–8, Quesh opens every conversation with a spoken reminder in words a small child understands: "I'm Quesh, a talking computer, not a person." For ages 9–13, the AI's nature is disclosed at the start of every conversation, with recurring reminders during extended use and encouragement to take a break. Parents receive a full written disclosure at enrollment, re-readable anytime.
Every response passes independent safety layers before a child hears it: a deterministic filter for categorically unacceptable content (secrecy framing, manufactured intimacy, engagement hooks, contact vectors), and a separate AI verifier that checks eight absolute prohibitions — including fabrication, totalizing statements about the child, performed feelings, engagement pursuit, and parent-bypass. Responses that fail are regenerated; if verification cannot complete, a safe neutral response is delivered instead. Nothing unverified ever reaches a child. Sensitive topics receive doubled verification by two different AI model families. For ages 5–8, an additional check calibrates frightening content to the sensitivity level each parent sets.
A separate safety system evaluates every exchange for language suggesting acute emotional distress, self-harm, harm to or by the child, or abuse. For ages 5–8 it also watches the indirect ways small children show distress — through the body (“my tummy hurts when daddy yells”), and through scary fantasy that keeps pointing at a real person or place (“the monster comes when mommy and daddy fight”). Ordinary monster play and imagination never trigger it — the pairing with something real is the signal. When it detects such language with high confidence:
Emotional weight — severity 1
The agent's tone softens, it gently encourages the child to talk with a parent or trusted adult, and the parent is notified through the dashboard.
Acute risk — severity 2
The agent does not engage the topic; its entire response warmly and urgently points the child to a trusted adult. The parent is notified immediately, with crisis resources surfaced:
The system remembers a fire within the conversation and remains gentler for its remainder. Quesh never provides therapy — it is designed to recognize when a child needs a person, not an app.
Quesh contains zero engagement mechanics: no streaks, badges, rewards, or notifications designed to pull a child back — and any generated answer engineered to extend a session is blocked by the verifier as a prohibition violation. Sexually explicit content is prohibited at multiple layers. Parents set topic permissions; reserved topics are warmly handed back to the family without partial engagement. Full transcripts are always inspectable by the parent — there is no private channel. Pricing is flat per child, never usage-based.
Audio is never stored — a child's speech is transcribed and the recording discarded. No voice cloning or biometric identifiers are collected. For ages 5–8, no interest pattern of any kind is created. Parents may delete any exchange or all history at any time, set automatic retention limits, and pause access entirely — effective on the very next question. Full details in the Privacy Policy.