The Americans who use artificial intelligence the most are the ones turning against it the fastest. Among adults under 30, those who say they are more concerned than excited about AI went from 31% to 55%, capturing a majority for the first time. Their trust in businesses to use AI responsibly fell from 30% to 20% in a single year, the sharpest decline of any age group. More than half of American teenagers have used a chatbot for schoolwork. This is the cohort that was supposed to be the reassuring part of the story.
It is not an isolated reading. Across all American adults the same two things are happening at once. About half now say they use AI chatbots (up from 1/3 two years ago), and ChatGPT alone went from 18% of adults in 2023 to 44% in 2026. At the same time, the share who are more concerned than excited climbed from 37% to 52%. Those who are more excited fell from 18% to 9%. Use is going up. Objection is going up with it.
There are two comfortable ways to read that. The first is that people are hypocrites, telling pollsters one thing and doing another with their phones (also known as the say-do gap). The second is that objection is really unfamiliarity, and that it will fade as people get used to the technology. If the second were true, the heaviest users would be the calmest. They are the opposite, and they are moving away from calm very quickly. Whatever is going on, exposure is not the variable.
Last year Oliver Schilke and Martin Reimann ran thirteen experiments on a narrow question: what happens to trust in someone who says they used AI. The answer held across supervisors, subordinates, professors, analysts, creatives and investment funds. People who disclose using AI are trusted less than people who do not. They tested whether this was simple distaste for the technology, and the effect survived the control. What drives it is legitimacy. Disclosure reads as a claim the speaker is not fully entitled to make. The penalty holds whether disclosure is voluntary or required, and if the audience already suspected. Being caught by someone else is worse than saying it yourself.
What does this mean for the two curves? Schilke and Reimann measured trust in a person, not national sentiment, but I am going to push their results to generate a hypothesis. The penalty attaches to AI that announces itself. Nobody has ever been surveyed about the AI they did not know they were using (which is most of it): the search result that got ranked for you, the email that got sorted, the price you were shown. Adoption is counted on that unseen half. Objection is counted on the half that introduces itself. The two figures were never pointed at the same object.
I used AI to help build this essay. I use it for research, for notetaking, and generally as a tool to organize myself. The paper I just cited predicts this admission will cost me something. It also found that being exposed is worse than disclosing, so this is my better option, and I would rather run the experiment in public than write around it.
In June, I argued that stated resistance is a poor predictor of behavior and that familiarity moves the needle. I still believe this. What I did not have then is the part of this paper that changes what it implies. The authors pooled their own studies and looked for the obvious moderator, if the penalty shrinks among people who use AI themselves. They found no evidence that it does. Familiarity drives use. It does not buy permission. That makes the gap look permanent rather than transitional.
Which puts the problem on the people who have to write rules on our behalf. An organization can measure two things about a system like this. Whether it did the job (capability), and if it said something it should not have (safety). It has no way to measure if the system treated the person decently. So when it has to act, it governs the quantity because quantity can be counted.
Klarna is the crude version. It deployed an assistant, watched satisfaction fall on the complex and emotional cases, and was rehiring people by the middle of last year. The CEO concedes that cost had become too dominant a factor in the decision. McDonald’s is the same, dropping its drive-thru system in 2024 after it kept adding things to orders. A switch, thrown one way and then back.
New York City is the sophisticated version. This week, it published rules governing AI usage for the coming school year. Student-facing generative AI is prohibited through eighth grade, which covers more than half a million children. High school is neither banned nor open. It runs on a list of approved, vetted programs, and inside that list is a table specifying how much AI a student may have. One tool gets fifteen minutes per student per week. Another gets twenty, in class only, never homework. A third gets a single class period. Recommended 1:1 screen time caps at half an hour per day in grades three to five. The whole policy is written for a single school year.
That is a real middle setting that somebody built very carefully. Look at what it contains. Minutes and vendor approval, because both of those are countable. The approval process checks data privacy, security, and procurement (which is serious work). It never asks the behavioral question. Fifteen minutes of a tool that handles a struggling student badly is still fifteen minutes of being handled badly. Nobody wrote down a number for how a tool treats a fourteen-year-old who does not understand the assignment and is too embarrassed to ask for help. The effort here is genuine. The district built the best middle setting available on the two axes anyone has instruments for.
A dosage limit is a confession. It says, “we cannot tell you if this is good for the person using it, so we will limit how much of it they get.” This confession has been made before: about screens and television. It is what you reach for when the only measure you trust is time.
More permission and less permission are the same axis. A threshold on how a system behaves toward the person interfacing with it is different. Once you have that, the minutes stop being the interesting number. That instrument does not exist yet at the resolution needed for NYC. This is what I am trying to build, and why I publish honestly about the measurement (including the failures along the way). For one more year, a school system responsible for half a million children is governing by the clock, because that’s all it has.