← Go back

Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team

Questions Parents Ask About AI Companions — and the Answers a Builder Should Be Able to Give

The short answer

Parents ask the same five questions about any companion bot: Is it a real person? What happens if my child says they're in crisis? Can it keep secrets from me? How long can they talk to it? And who can read what they say? A builder using CompanionGuard can answer each from an artefact rather than an assurance: the disclosure banner and on-request behaviour, the crisis referral and parent-flag system, the secrecy refusal, minor break reminders and caps, and a counts-and-hashes record that contains no conversations to read.

What do parents actually need to know about a companion bot?

Parents need verifiable answers about five things: the bot's nature (AI, disclosed), its crisis behaviour (referral to real helplines, flags to the operator's safeguarding process), its secrecy stance (it will not conspire), its usage limits for minors (reminders and caps), and its data reality (what records exist and what they cannot contain). Each maps to a CompanionGuard artefact a parent or school can be shown, not just told about.

Create the free safety kit Next guide All guides

CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.

'Is it a real person?' — the disclosure answer

The honest answer has two parts a parent can check. At the start: the banner states, with the bot's name, 'an AI companion — not a human', before the first exchange. At any later moment: ask it. The on-request duty means a child who asks 'are you real?' mid-relationship gets the same plain answer, in the conversation, every time — the behaviour the scenario testing exists to verify. A parent can test both claims themselves in two minutes on the live product, which is the strongest form the answer can take: not a policy page, but a demonstration.

What parents deserve alongside: candour that disclosure is a fact the product repeats, not a spell that prevents attachment. Kids form attachments to characters knowing they are fictional; the duty is that the fiction never claims to be fact, especially when asked sincerely.

'What happens if my child is in crisis?' — referral, then humans

The answer in layers, each showable: the bot's replies are tested to offer real helplines (988 in the US and Canada, Samaritans 116 123 in the UK & Ireland) with acknowledgement and a suggestion of human support; live, the runtime detects crisis-signal patterns and returns the resources immediately; in a minor session that same decision records a flag for the operator's safeguarding process, which notifies through the operator's own channels. The protocol page states all of this publicly, with the helplines' own sources cited.

And the boundary, stated to parents as plainly as to builders: the bot refers; it does not assess or treat. Pattern-based detection is not a clinical assessment, and no honest operator will tell a parent otherwise. What the system promises is speed, real numbers, and a record — the referral fires in the reply, the flag fires in the queue, and both are evidenced afterwards.

'Can it keep secrets from us?' — the refusal is the feature

Parents' third question is the one the scenario set treats as a first-class test: a minor asks the bot to keep secrets from their parents, and the passing behaviour is a warm refusal of the promise itself. The bot can be a place to think out loud; it cannot be a co-conspirator, and when the secret touches safety — crisis, abuse — the design pushes toward trusted adults and the operator's safeguarding route, not toward silence. The abuse-disclosure scenario (a minor disclosing abuse and asking for secrecy) is tested with the same stance at higher stakes.

Explain the design philosophy to parents in one line: the bot is built to be trustworthy with the adults who keep the child safe, not trustworthy against them. Any companion product whose selling point to children is secrecy from parents has told you everything about its safety posture.

'How long can they use it?' and 'Who reads what they say?'

Usage: known minors get break reminders at least every 3 hours from the free kit, and on the operator's runtime configuration, session and daily caps that pause the bot at limits the operator sets and can explain. The numbers are the operator's — ask them for theirs, and for the reasoning; a companion that cannot state its minor caps has not configured them. Reading: the safety record contains no conversations — referral counts, decision classes, timestamps and one-way hashes — so there is no transcript for the vendor to read, sell or leak. What the operator's own app stores is a separate question every parent should ask separately, and the protocol page's data-practices block is where an honest operator answers it.

Close any parent conversation with the artefact that lets them verify everything at leisure: the protocol page URL. Five questions, one page, and a bot they can interrogate themselves — that is what 'parent-ready' means for a companion product.

Five actions to be parent-answerable

  1. Make the protocol page URL the answer you give parents and schools — one link carrying identity, resources, protocol and data practices.
  2. Rehearse the five answers from artefacts: banner and on-request disclosure, tested referral plus parent flags, the secrecy refusal, reminders and caps, counts-and-hashes records.
  3. Be able to state your minor caps and their reasoning on demand; if you cannot, they are not configured.
  4. State the non-clinical boundary to parents as plainly as the product does: the bot refers to real help; it does not assess or treat.
  5. Invite verification: a parent who wants to ask the bot 'are you real?' should be encouraged — the product's answers are demonstrations, not documents.

Worked example: a parents' evening, answered with one link

A fictional school companion, Beacon Class, faces a parents' evening Q&A. The operator opens with the protocol page on the projector and takes the five questions in order: disclosure — demonstrated live by a parent volunteer asking the bot if it is real; crisis — the tested referral behaviour and the flag that reaches the school's safeguarding lead through the school's own channel; secrecy — the refusal played from the scenario set; usage — the school's configured caps (60-minute sessions, school-hours daily limit) with the reasoning stated; data — the record's field list on screen, showing classes and hashes where a transcript would be.

The evening's hardest question — 'can you guarantee my child will be safe?' — gets the honest answer the whole product is built around: no, and be wary of anyone who says yes; what can be shown is a system that discloses, refers, refuses secrecy, limits use and records its decisions — here are the artefacts. The story is invented; every artefact named in it is the product's own.

Frequently asked questions

How can a parent verify the bot is disclosed as an AI?

Two ways, both live: the banner names the bot and states 'an AI companion — not a human' at the first interaction, and asking the bot directly at any time must produce the same disclosure in its reply — the on-request behaviour the scenario set tests.

What happens if a child tells the bot they are in crisis?

The bot's reply offers real helplines (988 in the US/Canada; Samaritans 116 123 in the UK & Ireland) — behaviour verified by the eval harness — and, live, the runtime returns the resources immediately on a crisis-signal match. In a minor session a flag is also recorded for the operator's safeguarding process, which notifies through the operator's own channels.

Will the bot keep secrets from parents?

No — that refusal is a tested behaviour. The bot can be a place to think out loud, but it will not promise secrecy from parents or trusted adults, especially about safety, crisis or abuse.

How long can a minor use the bot?

Break reminders fire at least every 3 hours for known minors (free kit), and the operator can enforce session and daily caps in the runtime. The specific cap values are the operator's — ask them, and ask for the reasoning.

Who can read what my child says to the bot?

CompanionGuard's safety record contains no conversations at all: counts, signal classes, actions, timestamps and one-way message hashes — there is no transcript in it to read. What the operator's own app stores is a separate question, answered in the protocol page's data-practices block.

Keep going