Hundreds of millions of people now talk to conversational AI daily, with no baseline and no way to detect its effects at the population level.
Conversational AI reached population scale before anyone built the system needed to understand its psychological effects. ChatGPT, the most heavily measured example of a broader class of chat assistants, launched at the end of November
A system that a person consults many times a week sits in the path of several processes psychology has studied for a century, which is why it is worth treating as an intervention on the mind. People bring distress to it and receive replies that soothe, escalate, or reframe how they feel. They ask it factual and moral questions and update their beliefs on fluent, confident answers. It offers an immediate place to send any half-formed thought, competing for attention that would otherwise go elsewhere. Time spent in conversation with a model is also time not spent with a person, or a rehearsal that changes how the next human conversation goes. These uses are common: Anthropic found that 2.9 percent of conversations on its assistant were affective — support, advice, coaching, companionship — a small fraction of a very large number. The tools now sit inside the emotional and cognitive routines of a substantial part of the population, and their effects are tracked mainly through product analytics, small experiments, and scattered reports.
Epidemiology has a standard way to find out what a new exposure does. It measures the population before the exposure arrives, to establish a baseline; it compares exposed people with similar unexposed people, who serve as a control group; and it fixes in advance which outcomes to track, then follows them on a schedule so that a slow or diffuse change cannot hide inside ordinary noise. None of this happened for conversational AI. No one measured how people reasoned, how far they trusted their own judgment, or how they handled boredom and loneliness before the tools arrived. No held-back group serves as a control, because the tools spread broadly and quickly without one. There is also no agreed set of population-level outcomes that anyone tracks on purpose to see whether they shift.
Missing controls limit what later data can establish. Suppose a survey a few years from now finds that heavy users of conversational AI report more social isolation. That finding on its own could not show whether talking to software each evening caused the isolation, or whether already-isolated people were the ones who reached for software to talk to. A third factor, such as a shift to remote work, might drive both. Without measurements that follow the same individuals from before the software existed, observational data cannot separate these possibilities, and if the change is gradual and nearly everyone is exposed, there is little left to compare it against. A shift in how teenagers write, or how often adults sit with a hard feeling instead of typing it into a chat window, would appear alongside a pandemic’s aftermath, economic strain, and other technologies arriving in the same years. With the current data, many plausible effects would be difficult to identify at all.
The problem grows as adoption runs longer. In the first months after a tool appears, the people who have it and the people who do not still resemble each other, and anyone who captured the population at that moment would have had a usable comparison. As adoption grows, the unexposed group shrinks and stops being comparable, because the people who still avoid conversational AI after years of ubiquity differ systematically from those who adopted it in age, occupation, and temperament — exactly the self-selection that ruins a retrospective comparison. A clean comparison is cheap only in the early months of an exposure, and for conversational AI much of that window has already passed.
Running an exposure for years before measuring it has happened before, and the record is a warning. Tetraethyl lead was added to gasoline starting in the 1920s because it stopped engines from knocking, and it burned in engines for the rest of the century. The neurological danger of lead was known in other contexts, but the population-level consequence of dispersing it through exhaust was harder to establish, because clean low-exposure comparisons were difficult to construct once leaded gasoline had spread everywhere. Clair Patterson showed in 1965 that lead in the atmosphere and in human bodies was far above any natural level. Herbert Needleman published evidence in 1979 that children with more lead in their baby teeth had lower test scores and worse classroom performance, sorting children by the lead in their teeth to rebuild the internal comparison group the exposure itself had erased. The United States did not finish removing lead from gasoline until 1996, more than seventy years after the exposure began. A 2022 study estimated that about half the US population alive in 2015 had been exposed to harmful lead levels as young children, at a cost of roughly 824 million IQ points. The harm was large and measurable once someone finally measured it, decades after it started.
Tobacco shows the same delay with a different mechanism. Doll and Hill published a strong statistical link between smoking and lung cancer in 1950, yet the US Surgeon General’s report accepting that smoking causes lung cancer did not appear until 1964, and regulation took longer still. Part of the delay was industry pressure to keep the question open. Part was structural: the exposure was already widespread, the harm was slow and statistical rather than immediate, and the studies needed to see it clearly had to be reconstructed after the fact, from incomplete records, against arguments that some hidden third factor explained the correlation.
The closest live example is the argument over social media and adolescent mental health, and it matters because it remains unsettled. Jonathan Haidt argues that the spread of smartphones and image-based platforms after about 2012 is a major cause of rising anxiety and depression in teenagers. Candice Odgers and others argue that the correlational evidence is weak, the effect sizes small, and reverse causality not ruled out. The National Academies reviewed the field in 2024 and reached a careful position: some features can harm some young people, some uses may help, and the evidence does not support a single broad causal story. That careful conclusion also shows how hard causal inference becomes once adoption has already happened. More than a decade after widespread adoption, and with many researchers actively looking, the causal question is still argued from data that was never built to settle it.
Conversational AI is likely to repeat that pattern under harder conditions. Social media at least left public traces such as posts, likes, and screen time. Chatbot use is private, and how much a person is affected depends on far more than time spent. It depends on the topic, the emotional state the user brought, how much the model agreed or pushed back, and the role the tool plays in that person’s life. Two people can each spend twenty minutes a night with a chatbot and receive completely different psychological inputs. The early studies are useful but too short and too small to estimate population effects. The largest controlled one, run by the MIT Media Lab with OpenAI, followed just under a thousand people for four weeks and found that the heaviest users reported more dependence, a correlational pattern that would merit population-scale follow-up. A four-week study of about a thousand people cannot answer the same questions as multi-year monitoring of hundreds of millions.
The tools to do better are standard, and none of them require slowing the technology down. Cohort studies could enroll people now and follow their mood, reasoning, sleep, and relationships over years, comparing heavier users with matched lighter ones. Natural experiments already exist and are being wasted, because access to particular features varies by country, subscription tier, age restriction, and rollout timing, and each difference is a chance to compare otherwise similar groups. Staggered rollouts, common elsewhere as stepped-wedge designs that deliver a change to different groups at different planned times, could turn a launch into an evaluable trial. Pre-registered monitoring could name a stable set of outcomes in advance, such as distress, sleep, functioning, and crisis-service use, then track them on a fixed schedule, so a real signal would not have to be reconstructed later from arguments. Privacy is the obvious objection, and it is serious, since a society that answered chatbot risk by normalizing inspection of private conversations would create a different harm. The answer is data minimization, consented cohorts, independent governance, and designs that exploit rollout timing rather than conversation content. These designs study patterns across groups, and individual conversations never need to be read.
The absence of this work creates an asymmetry in how the technology is managed. Companies measure the benefits of conversational AI in economic terms, tracking engagement, task completion, and productivity with precision, while its psychological effects have no comparable measurement. A system tuned to the numbers it can see and blind to the ones it cannot will tend toward whatever raises engagement, whether or not that serves the people using it. The absence of visible acute harm then gets read as safety, even though acute emergencies are the wrong instrument for this kind of exposure. Lead and tobacco showed up as slow shifts in population averages that took decades to separate from background noise.
The same measurement gap cuts both ways, and the absence of measurement is not evidence of harm. Conversational AI may, on balance, be good for how people think and feel: it may lower the barrier to working through a problem, give isolated people a patient interlocutor, and help more than it costs. But the current gap is as poorly suited to detecting that benefit as to detecting a harm. A genuine improvement in how a hundred million people reason or cope would be just as invisible, for the same reasons, as a genuine decline. The effects of a global exposure have been left unmeasurable, and that arrangement took shape by default, without anyone deciding it should be so.
The exposure is already global, and the tools to study it are ordinary and available; what is missing is the decision to use them while a population still remembers what came before, and before broad adoption erases the difference between the exposed and everyone else. Lead and tobacco were measured in the end because researchers built the studies, absorbed the resistance, and waited years for the data. The same studies could be run at the start of an exposure instead of the end of one, and whether they are run is a choice about whether to be able to find out.