Half of America uses AI chatbots. That is not the same as wanting them.

Pew's 2026 survey shows US adult chatbot use has climbed from 33% to 49% in two years, while 63% of adults say AI is moving too quickly and 71% expect it to make their personal data less secure.¹ The standard reading of numbers like these is that adoption settles the argument. I do not think it does. What the survey actually documents is a large and growing population using tools they distrust, and that is a different story with different implications.

The revealed-preference argument, and why it does not hold here. The instinct in enterprise rollout decks and in a lot of tech commentary is to treat usage as a proxy for consent. People are using the thing; therefore they want the thing; therefore the concerns are overstated. This is the move Pew's data quietly breaks. Adoption nearly doubled at the same time as scepticism stayed put or hardened. The two curves are not converging.

Consider what would have to be true for the revealed-preference reading to work. Users would need meaningful alternatives, adequate information about what the tools do with their inputs, and the ability to opt out without professional or social cost. On all three, the actual conditions of chatbot adoption in 2026 fall short. ChatGPT alone is used by 44% of US adults.¹ Employers are integrating these tools into workflows. Search engines are surfacing chatbot answers by default. This is not a market of shoppers picking a preferred product. It is a shift in the default substrate of everyday information work.

What "too fast" actually names. The 63% figure is easy to read as vague anxiety. I think it is more specific than that. When people say a technology is advancing too quickly, they are usually saying something about the gap between deployment and governance, between what the tools can do and what anyone has been asked about them. Pew's own finding that majorities lack confidence in either government or companies to regulate AI responsibly points the same way.² The concern is not the pace of capability. The concern is the pace at which capability is being pushed into daily life without a corresponding conversation about terms.

That 71% is the number I keep returning to. It is not a small dissenting minority worried about a distant risk. It is a supermajority of the adult population using a technology they expect to erode their own data security. If a food product had those numbers, we would call it a public-health story. Because it is software, we call it adoption.

The generational inversion. The pitch decks lean hard on a generational adoption story. Younger users will normalise these tools; scepticism is a boomer artefact; wait it out. Pew breaks this cleanly. Adults aged 18 to 29 are simultaneously the heaviest users of chatbots and among the most apprehensive about AI's societal impacts.¹ The people who use the tools most are not the people most reconciled to them. They are the people with the most exposure and, plausibly, the most texture in their view of what the tools actually do.

This matters because "the young will love it" has been doing quiet work in a lot of deployment arguments. Remove it, and you are left with the actual finding: familiarity is not producing acceptance. It is producing informed unease.

Who is exposed, and to what. Pew's racial and ethnic breakdown is worth sitting with. Asian adults use chatbots at roughly 7-in-10; Black, Hispanic and White adults sit around half.¹ The report's methodology should be read before drawing sharp inferences about why. But the exposure implication does not depend on the cause. Higher usage means higher exposure to whatever the tools embed — data extraction, algorithmic mediation, decisions made or suggested by systems whose training data and error patterns the user cannot inspect. If those embedded features carry real costs, and 71% of adults expect at least the data-security ones to, then the costs are not falling evenly.

I want to take the strongest counter-argument seriously. It runs like this: people routinely use tools they distrust when the immediate utility outweighs the diffuse risk. Social media is the obvious analogue. Cars, credit cards, most of consumer software. Perhaps "I use it and I think it is bad" is a stable and rational equilibrium, not a problem to be solved. Perhaps the survey is registering a background hum of tech-scepticism that never translates into anything.

The social-media parallel is where this argument is strongest and also where I find it most concerning. That equilibrium held for roughly fifteen years. The harms it produced — to attention, to teenagers, to the information environment, to the ability of institutions to hold shared facts — are now widely acknowledged, mostly after the window for structural intervention had closed. "Attitudinal scepticism is decoupled from policy pressure" is not a reassurance. It is a description of how a previous generation of tools escaped scrutiny while people expressed exactly the concerns Pew is documenting now.

What the survey is really telling deployers. For the companies rolling out AI to employees and customers, the Ciscos and HPs and state governments cited alongside these findings³, the honest reading is uncomfortable. Your users are not the enthusiastic early adopters of the internal pitch deck. They are, in majority, people who expect the technology to make their data less safe and think it is moving too fast. They are using it because you are deploying it, or because their peers are, or because the search box now answers them differently. That is a legitimate business condition to work with. It is not the same as demand.

The distinction matters because it changes what a responsible deployment looks like. If you assume adoption equals endorsement, you optimise for uptake and treat concerns as friction. If you accept that adoption is running ahead of consent, you have to build the consent back in — through defaults, through disclosure, through genuine opt-outs, through governance structures that give the sceptical majority a voice they currently do not have.

Half of US adults use AI chatbots. Nearly two-thirds think AI is moving too quickly. Seven in ten expect it to erode their data security. These are not contradictions to be resolved by picking one number to believe. They are the same picture: a technology being absorbed into daily life faster than the people absorbing it have been asked whether they want it. The interesting question is not whether that gap closes. It is who gets to decide when it does.

Footnotes

¶