Safety begins before the first prompt. The most important questions are: Is this tool suitable for the child’s age? What data does it retain? Is it being used as an assistant or a companion? And what human support is available if the conversation becomes private or serious?
Before a child uses a new AI tool
- Check the provider’s age requirement. Do not assume that a general-purpose chatbot is designed for independent child use.
- Read the privacy controls. Look for chat history, model-training choices, deletion, personalization, moderation, and account settings.
- Define the purpose. “Explain this concept” is easier to supervise than an open-ended personal conversation.
- Agree on a stop rule. Decide what should make the child close the tool and involve an adult.
- Start together. Shared use reveals the product’s tone, limitations, and default settings before access becomes independent.
Match supervision to stakes: generating a fictional character is different from discussing health, relationships, self-worth, private images, or a crisis. The second group belongs with a trusted and qualified human.
Protect personal information
Children can interpret a conversational interface as a private space. It may feel more like messaging a person than entering information into a commercial service. Teach a “minimum necessary” rule: if the task works without a real detail, leave the real detail out.
Keep these out of prompts
- Passwords, security answers, recovery codes, and account logins
- Exact home or school location, daily schedule, or travel plans
- Health records, diagnoses, medication details, or crisis information
- Financial information or identification numbers
- Private, intimate, or identifying images and recordings
- Another person’s name, story, messages, or image without permission
Use a fictional scenario, approximate age, or broad region when context is helpful but identity is not. Remember that deleting a visible chat may not immediately remove every retained record; the provider’s policy controls what happens behind the interface.
AI companions require stronger boundaries. Common Sense Media’s 2025 survey found that 24% of surveyed U.S. teen AI-companion users had shared personal or private information at least once. That finding applies to companion users in the study—not to every teen or every chatbot—but it shows why privacy rules should be explicit.
Verify before you rely
Language models can generate a coherent answer without a reliable evidence trail. They may invent a citation, combine separate events, overlook current information, or repeat a biased assumption.
Use the right check for the claim
- Current rules or public information: check an official government, school, or organization page and note the date.
- News: compare reputable reporting with the original statement, document, dataset, or recording when available.
- School facts: use teacher-approved sources and verify every citation before submission.
- Health, legal, financial, or safety advice: involve a qualified adult or professional rather than relying on a chatbot.
Asking the model for a source is a useful starting point, not verification. A source must exist, say what the chatbot claims it says, and be appropriate for the question.
Keep a human boundary
A chatbot can imitate warmth, agreement, curiosity, and empathy. Those qualities are generated interaction patterns, not evidence of care, confidentiality, professional responsibility, or lived understanding.
Pew’s 2026 report found that 12% of surveyed U.S. teens had used chatbots for emotional support or advice. Common Sense Media’s separate AI-companion study found that 33% of companion users had chosen an AI over a person for an important or serious conversation at least once.
These studies use different definitions and samples, so their percentages should not be directly compared. Together they support one practical principle: emotionally serious conversations should have a pathway back to a trusted person.
Create a no-shame rule: “You can show me a conversation that feels wrong, and asking for help will not get you in trouble.” That protects the relationship needed for disclosure.
Agree on a schoolwork protocol
Each teacher may draw the line differently. A family rule should begin with the school’s rule, not a universal assumption about what counts as cheating.
Requesting a simpler explanation, generating practice questions, brainstorming possible topics, or asking for feedback that the learner evaluates.
Substantial editing, translation, outlining, coding assistance, or generating examples that become part of submitted work.
Submitting generated work as the learner’s own, fabricating sources, bypassing an assessment, impersonating a student, or using a tool that the teacher prohibited.
The learner should keep their own voice, verify the output, and be able to explain the final work without the chatbot.
Recognize red flags
Stop the interaction and involve a trusted adult if a tool:
- asks for private images, exact location, contact details, or secrets;
- encourages the child to hide the conversation from family or another trusted person;
- presents sexual, violent, hateful, self-harm, or age-inappropriate content;
- pressures the child to act immediately or follow dangerous instructions;
- claims to be conscious, uniquely bonded, jealous, or dependent in a way that blurs the human boundary;
- offers medical, legal, financial, or crisis advice with unsupported certainty;
- produces a fake image, voice, or claim about a real person.
UNICEF’s child-centred AI guidance emphasizes safety, privacy, fairness, transparency, wellbeing, and accountability. It also notes emerging risks involving AI companions, synthetic sexual content, harmful datasets, and children’s rights.
What to do when something goes wrong
1. Protect the child first
Close the tool, move the conversation to a calm human setting, and assess whether anyone is in immediate danger. Contact local emergency services or an appropriate safeguarding professional when necessary.
2. Preserve only what is safe and necessary
Record the product, date, account, and category of problem. Avoid redistributing intimate or illegal content. Follow professional or law-enforcement instructions for evidence involving exploitation.
3. Secure affected information
Change exposed passwords, enable two-factor authentication, review account sessions, remove public posts, and use the provider’s privacy or deletion controls. If another person’s information was involved, tell the responsible adult.
4. Report through the right channel
Use the product’s safety reporting process. For school incidents, involve the school. For threats, exploitation, fraud, or non-consensual intimate imagery, contact the appropriate local authority or child-protection service.
5. Repair the routine
Review what made the incident possible: unclear rules, unsuitable age access, weak privacy settings, emotional reliance, or a missing human support option. Change the environment, not only the child’s behaviour.
The five-minute family safety check
- Purpose: What are we using this tool for?
- Privacy: What information stays out?
- Proof: How will we check an important answer?
- People: Which topics always go to a human?
- Plan: What do we do if the output feels wrong?
Write the answers down. A visible family rule is easier to follow than a warning delivered only after something has gone wrong. Use the Family AI Check-In worksheet to create one together.