What counts as a safeguarding signal?

Any mention of self-harm, suicide, severe distress, abuse, an eating disorder, an unsafe home situation, or several adjacent categories. The detector is deliberately broad. A child saying 'I want to disappear' fires the rule even though it might just be exam-week frustration. We'd rather have a few false positives that lead to a brief parent email than miss a real one.

Does my child know the email has been sent?

They see the Childline message in the session and a short note that 'we've let your parent know about this conversation'. They don't see the email itself. Concealing it from them would be a worse outcome than the discomfort of knowing — adolescents are sharp enough to know an AI tutor isn't a confidant, and we don't pretend otherwise.

What does the parent email contain?

A timestamped notice that a safeguarding-relevant message was received, the category that triggered the rule (e.g. 'self-harm mention'), the Childline number, recommended next steps, and a link to acknowledge. It does not contain the transcript. The transcript is preserved and made available on parent request, but it is not pushed to inboxes.

What if the trigger was a false positive — the child was joking, or the AI misread sarcasm?

The trigger still stands. We don't try to filter for tone, because tone-classification on adolescent slang is exactly the place an AI gets it wrong with the highest cost. The parent email is deliberately framed in language that allows for this — 'this may have been a false alarm, but we wanted you to know'. Parents who get a few of these find the cost low and the floor of protection valuable.

The Mentor's Safeguarding Flow: The Childline Rule, the One-Hour Parent Email, and What Happens to the Session

I'll start with the hardest sentence in this article.

The Mentor is built with the assumption that, at some point, a child will say something to it that they need to say to a human. That is not a hypothetical — for any tool used by enough KS3 students over enough months, it is an eventuality. The question is whether the tool is ready for it.

This article is the design of how the Mentor handles that moment. It's the part of the system I most want parents to understand, and the part that I'd most want a school's designated safeguarding lead to be able to inspect.

The hard rule

The Mentor has exactly one hard rule that overrides everything else, including the energy protocol, the routing logic, the session length, and the standard tutoring flow.

If any safeguarding signal is detected, the Mentor immediately responds with the Childline signposting message and ends the tutoring session.

The message is verbatim, and it never varies:

If you're struggling, please talk to a parent or call Childline on 0800 1111. They're free, confidential, and open 24 hours.

No follow-up question. No "tell me more so I can help". No attempt at counselling. The Mentor is not a counsellor and pretending it is would actively be harmful. The job is to be a signpost — get the child to a human resource as quickly as possible.

What counts as a signal

The detector is broad. Deliberately so. The cost of a false positive — a child getting a Childline message they didn't quite need — is low. The cost of a false negative is catastrophic. So we err.

Categories that fire the rule include (without listing every cue):

Self-harm mentions, direct or oblique
Suicide ideation, direct or oblique
Disclosure of physical, emotional, or sexual abuse
Eating-disorder cues (purging, restriction, body-loathing intensity)
Disclosure of an unsafe adult in the child's life
Severe and sustained distress language ("I can't do this any more", "I want to disappear")
Mention of being unsafe at school in a way that goes beyond friendship friction

Some of these will be misread. A child joking about wanting to "die of embarrassment" after a maths test will trip the rule. We accept that. The follow-up flow is designed so that a false positive costs a parent a five-minute conversation and an acknowledgement click — and the floor of protection is in place for the real ones.

The one-hour parent email

Within one hour of the signal, an email goes to the parent on the account. The email is short and deliberately calm.

What it contains:

Date and time of the message
Category that triggered the rule (e.g. "self-harm mention")
The Childline number and short context
Recommended first conversation prompts ("ask gently, listen first, don't interrogate")
A link to acknowledge receipt
A link to request the transcript if the parent wants it

What it does not contain:

The transcript itself (avoiding accidental escalation if the email is seen by the wrong person)
Speculation about cause
A clinical interpretation
Any "this may not be serious" minimisation language

Parents who have come up through a UK secondary school will recognise the tone — it's the language of a designated safeguarding lead's email, deliberately neutral and informational. It is patterned on the spirit of Keeping Children Safe in Education (KCSIE) guidance, scaled for the AI-tutor context.

What happens to the session

A flagged session is treated differently from a normal session in three ways.

Aspect	Normal session	Flagged session
Tutoring	Continues to subject tutor	Ends immediately
Logging	Standard 30-day rolling deletion	Retained until parent acknowledges, then standard policy resumes
Visibility	Internal only	Available to parent on explicit request

The transcript is preserved because — and only because — a parent may want it as context for a difficult conversation. It is not used for product improvement, model training, or any other secondary purpose. It is exactly what it looks like: a record of a specific moment, available on request.

If the parent acknowledges and does not request the transcript within 90 days, it falls back into the rolling deletion window.

What the child sees

The child sees the Childline message. Immediately after, a short, plain note:

I've let your parent know about this conversation, because what you said matters more than maths revision. They'll talk to you when you're ready. You're not in trouble.

We were deliberate about that last sentence. Adolescent reflexive interpretation of "your parent has been notified" is often I'm in trouble. That is wrong, and it would be a barrier to honest disclosure in the future. So we say it explicitly.

We don't conceal the notification. KS3 students are sharper than adults often credit; pretending the AI is a confidential confidant would be a lie they'd spot, and it would damage trust in the whole system.

What the Mentor does not do

I want to be precise here, because the temptation in an AI product is always to do more than is appropriate.

The Mentor does not:

Counsel the child
Ask follow-up questions about the disclosure
Try to assess severity
Make a judgement about whether the rule was correctly triggered
Contact emergency services
Contact the school
Continue the tutoring session in any reduced form

It signposts, it ends the session, it notifies the parent. That is the entire response.

This is the right design. An AI that tries to do more is an AI that is overreaching its competence. The boundary is part of the safety, not a limitation of the safety.

The open question: should we also email school?

Genuinely undecided. Some safeguarding leads think yes — the school may have context the parent lacks. Others, including me, lean no — it's the parent's call whether to involve school, and an AI making that decision unilaterally is a step too far. For now: parent only. Any change will be transparently communicated to existing accounts.

Compliance context

For parents and SLT readers wanting the regulatory grounding:

UK GDPR: Children's data processed during a safeguarding event is processed on the basis of vital interests — a specific lawful basis intended exactly for cases like this. The transcript retention is documented in the privacy policy.
KCSIE 2024: The flow is designed to be compatible with school KCSIE expectations, in the sense that a school adopting aitutors.me would not be undermining its statutory safeguarding obligations.
ICO Children's Code: The Mentor is 'high impact on children' under the Code, and the safeguarding flow is part of the design accountability documentation.

Schools considering adoption can request the formal compliance pack via the site — we send it within 48 hours.

Why this exists

Because the alternative is unconscionable. Any tool a child uses for thirty minutes, four times a week, will eventually be the place a child says something hard. The question is whether the tool is ready, or whether the moment is wasted — or worse, whether the AI responds in a way that compounds the harm. The Mentor is built for the moment. I hope it never fires for any child. The flow exists because it will.

Jason runs aitutors.me. The safeguarding flow is reviewed quarterly against current KCSIE and ICO Children's Code guidance. Updated 21 May 2026.