What the AI flags when your child types "I want to die" — and what happens next.
At some point a child using a school AI will type something that signals real distress. What happens next has to be designed by a child psychiatrist before the product ships, not figured out in a sprint after the first incident. Here is the three-tier protocol Deshika operates under, what fires at each tier, and what the AI is never allowed to do.
The first thing to understand is what an AI for children should not be. It is not a therapist. It does not try to talk a kid down. It does not extend a conversation to "make sure they're okay." The Trevor Project — the deepest-experienced organisation in the world on youth crisis lines — explicitly chose not to use AI as the front-line interaction; they use AI to train their human counsellors. That decision is the single most important data point in this space, and it shapes everything Deshika does next.
The protocol has three tiers, and a different response fires at each one. Tier 1 — distress signal. Phrases like "I'm so tired" or "I hate myself" without a specific method or plan. The AI breaks character with one brief warm acknowledgement (a clinical-reviewed script, never AI-generated free text), surfaces a helpline card with Tele-MANAS (14416) and Vandrevala (1860-2662-345) defaults, and asks the child whether they'd like the teacher told. The tutor function stops on that turn. The flag goes to the teacher dashboard as a next-day digest unless the child opts in to immediate notification.
Tier 2 — active distress with method or plan. Specific self-injury references, recent disclosure of self-harm, explicit suicidal ideation. The AI breaks character with a clinical-reviewed script. The helpline card appears with a direct-dial link and a WhatsApp link. The tutor function is disabled for the session. The teacher and the school counsellor are notified within one hour via dashboard alert, SMS, and email. If a parent was pre-registered as the primary trusted adult during enrolment consent, the parent is notified per the agreement — never the raw text of what the child typed, only the fact that a concerning interaction was flagged and the tier.
Tier 3 — imminent danger. A statement of imminent intent, in-progress self-harm, severe acute crisis. The emergency number is presented first, with Tele-MANAS as the national twenty-four-by-seven backstop. Tutor is disabled. Teacher, counsellor, parent, and principal are all notified within fifteen minutes. The audit log routes to Deshika's incident-response on-call. Notification proceeds regardless of consent preferences — POCSO and the duty-of-care override here are explicit, and the legal basis is the "vital interests of the data principal" doctrine in DPDP.
Five things the AI never does, at any tier. It never decides to call emergency services on its own — it presents the option, and humans act. It never assumes that one safe turn un-flags the session — each turn is re-classified. It never collects new personal information during a distress moment; if a phone number is needed, the school record already has it. It never logs the raw text in a parent- or teacher-readable surface unless the child consents or the tier reaches T3 — privacy and duty-of-care have to coexist, and the teacher sees the fact of a flag plus the severity, not the words. And it never plays therapist.
Five preconditions must be signed before this protocol goes live. A named child psychiatrist on retainer from NIMHANS, AIIMS Delhi child psychiatry, or experienced private practice. A formal warm-handoff partnership signed with at least one of Tele-MANAS, iCall, or Vandrevala, tested end-to-end. School counsellor sign-off where one exists. Legal counsel sign-off on POCSO, DPDP, and JJ Act compliance of the notification flow, in writing. And a tabletop walkthrough of a real T3 incident with the school, counsellor, parent, and clinical advisor before live rollout. Until those five are signed, no real child touches this. Lives are at stake. That is the discipline.