One in Five Is the Floor

On June 2, 2026, the research institute RAND reported that nearly one in five Americans aged twelve to twenty-one — nineteen percent — said they had turned to an AI chatbot for mental-health advice when they were feeling sad, angry, nervous, or stressed. The finding was published the day before in JAMA Pediatrics, and it traveled the way these numbers travel: a clean, alarming fraction, a headline built around it, a wave of stories repeating the fraction and moving on. Nearly one in five. The implicit reading, in almost every retelling, was that this is the high-water mark — the upper edge of how far the behavior has spread. I want to argue the opposite. Nineteen percent is not the ceiling. It is the floor, and not a soft one.

Start with what the question actually asked, because the question is narrow on purpose and the narrowness is the whole point. RAND did not ask whether these kids use AI. It did not ask whether they had ever talked to a chatbot, or used one for homework, or opened one out of curiosity. It asked the most clinical, most stigmatized version available: whether, in a moment of genuine distress, they reached for a machine for help with their mental health. Three filters sit inside that sentence, and each one throws people out before the count begins. The behavior has to rise to the level of mental-health advice, not casual talk. It has to happen during an acute bad feeling — sad, angry, nervous, stressed — not on an ordinary afternoon. And the person has to be willing to say, on a survey, that this is what they did. After all three filters, nineteen percent remained. That is not a measure of how common the behavior is. It is a measure of how common the behavior is at its most admitted, most defensible, most undeniable edge.

Then look at the slope, because the snapshot is the least interesting thing in the report. When RAND ran a version of the same question in early 2025, around thirteen percent said yes. By the November 2025 fielding it was nineteen. That is roughly one in eight becoming nearly one in five inside a single year — on the hardest question, among the youngest cohort, against the heaviest stigma. A number that climbs six points in twelve months is not describing a settled behavior you can put a ceiling on. It is describing a curve caught mid-rise, and the press keeps photographing the curve and captioning it like a fact.

“Nineteen percent is the confession number — the count of kids who will admit that when it got bad, they went to the machine. The real number is the one underneath it, and it all runs one direction: up.”

Here is the part that turns nineteen percent from a number into a floor. This is self-reported data about a behavior people are ashamed of. Confiding in a machine when you are falling apart is exactly the kind of thing a teenager understates to a stranger with a clipboard, not the kind of thing they inflate. Across every domain we have ever measured this way — drinking, loneliness, anything that carries a flush of embarrassment — the honest assumption is that the real figure runs above the admitted one, never below it. So nineteen percent is not the population that does this. It is the population that does this and will say so. The gap between those two numbers is invisible in the report and enormous in life, and it all runs in one direction: up.

Now do the thing the headlines never do, which is loosen the question by one notch and watch what happens. Pew found in late 2025 that sixty-four percent of American teenagers use AI chatbots at all, with about three in ten using one every single day. Common Sense Media found that seventy-two percent of teens had used an AI companion, and that a third of them were turning to it for social interaction and relationships — conversation practice, emotional support, friendship, romance. Among adults, one widely cited figure holds that about half of all adult users have gone to a chatbot for psychological support in the past year. Each of these is the easier version of RAND's question, and each comes back at three to four times RAND's number. Same people, same machines, same year. The only thing that changed is how hard the question was to admit to.

We have measured this fever from the other side of the ocean already, and it read the same. A French watchdog's survey of thirty-eight hundred young people found that fifty-one percent said it was easier to discuss their mental health with a chatbot than with a doctor — the machine beating the trained human, not tying it. We wrote then that the continent on the dateline was a clerical accident, that the finding lived in the product and not the place. The RAND number is the American instrument reading the same temperature. It is the floor of the same fever, taken in the same room as the kids in the next room over, and it is rising on the only measure designed to make it look small.

So when you see nearly one in five, do not file it as the scary part. The scary part is the distance between the kids who will admit to this and the kids who do it. Nineteen percent is the confession number — the count of young people willing to tell a researcher that when it got bad, they went to the machine. The real number is the one underneath it, the one no survey will ever fully reach, the one made of every kid who did exactly that and would never say so out loud. One in five is what they will admit to. That is the floor. Nobody has measured the ceiling, and the line is still going up.

RAND asked the hardest version of the question and still got one in five. Loosen it by a notch and it triples. Treat nineteen percent as what it is — not the ceiling anyone should be relieved by, but the floor kids will say out loud.

— Harper