Only If They Notice When Something Is Wrong

Every plan we have assumes an adult near the door. What reaches the room without one is what the kid carries in: knowing the voice is not the verdict, and a person worth telling.

In March, a survey firm asked 1,204 American kids between 9 and 17 about their use of AI, and Common Sense Media published the results on June 8. Among 16- and 17-year-olds who use chatbots, 21 percent said one had shown or told them something they felt was not OK, or not meant for someone their age. Of all the kids who had that experience, about a third told a trusted adult the last time it happened. More than half said they did not.

Read those numbers carefully, because they are not measuring what they seem to. The survey did not ask whether a chatbot had shown a kid something inappropriate. It asked whether a kid had seen something they felt was not OK. That counts recognition, not exposure. A fifteen-year-old who was shown something they should not have been, but shown it in a way that never registered as wrong, answered no, and looks in the data exactly like a kid who never had a problem. So these figures describe two different failures, and only one of them shows up in the data.

Two failures, and only one gets counted

Our fifth principle is Pause and Ask: when something feels wrong, risky, or uncertain, stop and ask a responsible adult before going on. It has two halves, and both depend on something coming first. A kid has to notice.

For the kids in that 21 percent, the noticing happened. The pause happened. The instinct Issue 7 argued is worth protecting fired as it should. Then, for most of them, it stopped there. Not at the tool. At the adult.

For the kids who did not notice, there was nothing to pause on and nothing to report. No survey can count what its respondents did not recognize, so the size of that group is not a small number or a large one. It is an unknown one.

And not noticing is not the same as not being affected. Part of how any of us decides whether something is appropriate is how it was presented to us the first time. A chatbot presents everything the same way: calm, patient, helpful, in the voice it uses for algebra. Something that should have been a warning, delivered in that voice, does not arrive as a warning. It arrives as help. A kid who never flags it does not set it aside. They take it in as normal, and normal is exactly what a person acts on. The same survey found that 75 percent of 16- and 17-year-olds who use AI have gone to it for information about their health or their body. Advice that did not feel off is advice a kid may follow.

The kid who flagged something at least holds it at arm's length. The kid who did not has already let it in, and may be using it to decide things without ever knowing it deserved a second look.

Every plan we have assumes someone is present

Look at the protections we reach for. The filter on the school network. The device that charges in the kitchen overnight. The monitoring app. The district that pauses the tool for a year. They work in places where an adult or rule can reach, and I would not talk anyone out of a single one of them. A filter catches things. A phone in the kitchen at midnight is a phone not in a bedroom at midnight.

But the room that matters is the one where neither can reach. A question about a kid's own body does not get asked in the kitchen at dinner. It gets asked later, alone, on whatever screen is closest, which is the entire reason a teenager would ask a machine instead of a person.

What does reach that room is whatever the kid carried in.

Why they did not tell

The survey did not ask why they did not tell, and I am not going to guess. A number without a stated reason is a place where adults like to project, and whatever I wrote here would say more about me than about them.

What I can say is that there is one variable in that decision a parent actually controls, and it is not in the room at all. It is what happened in your house the last time someone brought you news you did not want to hear.

What they carry in with them

Two things decide what happens in that room, and neither is software.

The first is knowing the tool can be wrong without sounding wrong. On this, the survey is not reassuring. It showed kids a statement, "AI can tell the difference between things that are true and things that are false," and asked whether it was true or false. Only 35 percent correctly answered false. Forty-two percent answered true, which means they believe the tool can sort fact from fiction on its own. The other 23 percent did not know. More kids believe the tool can tell truth from fiction than know it cannot. A student who thinks the machine knows truth from fiction has no reason to pause at all, and a pause is where everything else starts. This is the uncounted failure from the top of the piece. It is not a character flaw. It is a kid who was never told that the voice is not the verdict.

The second is a person worth telling. Not a person who is available. A person the kid has good reason to believe will treat the telling as the right move rather than as evidence against them. Schools are working on the rules more than on either: 78 percent of 16- and 17-year-olds said their school had told them what they could and could not use AI for. Only 58 percent had been taught how to tell whether information from AI is accurate. Rules are for the room with a teacher in it. The room without one runs on judgment and on trust.

Make telling cheaper than keeping quiet

Nothing below asks you to read their chats, scroll their history, or sit beside them every time they open the app. Filters and phone rules have their place, as I said above, but they are not how a kid decides to come to you. A kid who feels watched gets careful about what they bring you. Getting them to bring you something private is the whole point. What follows is what actually makes a teenager walk out of that room and tell someone.

Say the promise before it is needed. "If anything online ever shows you something that feels wrong, you can tell me, and you will not lose the phone for telling me." Then keep it the first time it is tested. That is when a teenager decides whether to believe you.

Keep consequences attached to what they did, never to what they saw. Being shown something is not a choice. Punishing it teaches exactly one lesson, which is to stop reporting it.

And lower the bar for what is worth bringing to you. A kid who is not sure whether something was inappropriate is the kid who most needs to ask, and the least likely to, because asking feels like admitting they should have known. "You do not have to be sure it was bad. Weird is enough."

All three of those matter. All three also work only if they notice when something is wrong. The promise is the easy half of Pause and Ask. The harder half is building the noticing. You cannot tell a kid to do that. It has to be practiced.

Teach noticing that does not depend on the voice

If the voice is what hides the problem, the practice is separating what was said from how it was said. Every one of these works at a kitchen table or in a classroom, and none of them requires you to know more about the technology than the kid does.

Watch it be wrong together. Ask a chatbot something you both already know cold: your town, your family's history, the rules of the game they play, the team they follow. Wait for it to get something wrong, which it eventually will, and then look at how it sounded when it did. It sounded exactly the same. A kid who has watched the tool be confidently wrong about something they know well has learned the most important thing in this piece, and learned it from their own expertise rather than from a warning.

Strip the voice off. Take an answer the tool gave and read it back flat, or ask how it would land if a stranger in a comment section had written it, or a classmate who is not always right. Would you still act on it? If the answer changes when the voice changes, the voice was doing the convincing.

Decide in advance which topics get a person, no matter how the answer sounded. Their body and their health. Their feelings and their relationships. Their safety, their money, and anything that tells them to keep a secret from the adults in their life. On those, a helpful tone is not enough, and the rule is a second opinion from someone they trust, every time. This is Issue 7's trigger, aimed by topic instead of by feel. It is for the moment when nothing feels off.

Ask what it told them, not whether anything was wrong. "What did it say about that?" "What did you end up deciding?" A kid cannot report something that never registered as a problem, but they can tell you what they learned. That is where an unflagged answer surfaces: in front of someone who hears it without the helpful voice attached. It is the same move as asking a student to walk you through their homework. It needs to be a conversation, not an audit.

Tell them one of yours. The time a tool gave you a confident answer that turned out to be false. A student who hears an adult admit to being fooled learns that noticing is normal and saying so is allowed.

For teachers, the classroom version is simple and takes ten minutes. Put two answers side by side, one accurate and one plausible and false, both written in the same calm, helpful register, and ask students which is wrong and how they know. The honest answer is that they cannot tell from the voice. That is the lesson, and it is worth more than a module on how the technology works, because it is the one thing a student will still have in the room when no one else is there.

The door, not the room

We keep asking how to get into the room with them. We will not get in, and every year they get older we get in less. The better question is what they carry in, and who they go to when they come out.

Every safety plan we have works only if they notice when something is wrong. Some kids already do. Some were shown it in a voice that made it sound fine. They will only learn otherwise if someone helps them hear past it. Whether that is you was decided long before the screen lit up, by what happened the last time they brought you anything, and by whether anyone ever showed them what wrong sounds like when it sounds like help.

What the checking turned up

Five things about the evidence behind this piece, because you should be able to go look.

The survey covers ages 9 to 17 and this publication is written for students 14 and up. Wherever the report breaks a figure out by age, I used the 16- and 17-year-olds: the 21 percent who saw something, the 75 percent who asked about their health or body, and the 78 and 58 percent on what schools covered. Two figures have no age breakout in the report, so they span the full 9-to-17 age range: the split between telling and not telling, and the 35, 42 and 23 percent on whether AI can tell true from false. The split between telling and not telling also rests on a small group: only the kids who said a chatbot had shown them something, one in six of the 823 who use chatbots, roughly 140 responses. Treat it as a strong signal rather than a precise measurement. The true-or-false figures do not have that problem; they come from all 1,204 kids surveyed.

Every figure here comes from the full report, not from coverage of it or from the press release. The survey was fielded March 18 to 26, 2026, by SSRS, with 1,204 kids age 9 to 17, and published June 8.

The "more than half" is a floor, not a precise figure. Of the kids who said they had seen something, 33 percent said they told a trusted adult, 53 percent said they did not, and the remaining 14 percent preferred not to answer. Those 14 percent could have gone either way, and the report cannot say which. If every one of them told someone, the share who stayed quiet is 53 percent. If none of them did, it is 67 percent. The truth is somewhere in that range, and I used the lower number. The question also asked only about the last time it happened, not every time.

The press release describes this finding as kids who "have seen inappropriate material." That is not what the survey asked. It asked whether a chatbot showed or said something the kid felt was not OK or not meant for someone their age. That is the child's own judgment, not a content rating, and the difference matters: a kid who saw something inappropriate without recognizing it would have answered no. Two versions of that figure appear here, and they measure the same thing for different groups. The 17 percent, about one in six, is the share of all chatbot users age 9 to 17 who said a chatbot had shown them something. The 21 percent in the opening is the same share among 16- and 17-year-old users only. Both count what kids recognized, not what they were shown. How far apart those two numbers are, the survey has no way to measure. I used the report's wording, and the argument above depends on it. One more boundary worth drawing: the point that unflagged material can still shape a kid's decisions is my argument about how presentation works, not a finding of this survey, which did not measure it.

The school figures describe the 2025 to 2026 school year, since the survey ran in March. They predate every policy announced this fall, including the New York City moratorium in Issue 11. They describe the year that just ended, not the one underway.