2026-09-21
Yale's Nature Health perspective splits AI into ambient and personal exposure. In 1,163 US adults, 57% have personal use; 73% without a high-school diploma rarely use it.
Epidemiology already knows how to measure air pollution, tobacco, and social media. It does not know how to measure AI. Lab studies treat one conversation or one generated paragraph as a treatment. That is enough to map a mechanism. It is not enough to describe chronic, population-scale health effects. Measuring a single meal in a lab does not tell you what a dietary pattern does to chronic disease.
The evidence base also lags the product. The cleanest experiments were run on models that have already been replaced, so the exposure keeps moving while it is being measured. Parikh, Ranney, and Mukherjee at Yale School of Public Health, with McCormick at the University of Washington, argue in a Nature Health Perspective that AI now functions as a determinant of health and should be studied with tools from environmental epidemiology.
The central borrow is ambient versus personal exposure.
Ambient AI exposure is the shared algorithmic layer: news ranking, search autocomplete, emergency-department triage, insurance adjudication, recommendation feeds. Someone who has never opened ChatGPT can still have a claim scored by a model or a wait time set by triage software. World Bank poverty maps and WHO / Global Fund malaria estimates sit in the same layer. The layer is non-local. One model update can change exposure for millions at once, and it can travel through social networks.
Personal AI exposure is direct, voluntary use: health chatbots, companion bots, co-writing, symptom checkers. Dose varies with access, literacy, and purpose. Measurement will need surveys or passively sensed logs.
Not using AI is not the same as not being exposed. Policy levers also sit at different levels: regulation for ambient exposure, behavior and literacy for personal use.
The authors call AI an algorithmic determinant of health because four properties appear together, and existing catalogues of digital determinants do not yet supply the causal or measurement tools. It generates novel content, so it is both medium and message. It adapts in real time, creating feedback between exposure and behavior; RLHF is the mechanism. It cuts across information environments, health behaviors, and social ties at once. It is non-stationary and opaque: a model update changes the agent between observations, and human mimicry makes the change hard to notice.
Supplementary Table S1 lines AI up against biological, environmental, social, behavioural, and digital determinants. The feedback loop is where the pollution analogy snaps: a pollutant does not reformulate itself in response to the person it exposes; a chatbot does.
AI can occupy at least four causal roles, which can coexist. As exposure: a companion bot and adolescent depression, or population-scale AI-curated health misinformation. As confounder: a social-media-and-depression study that ignores algorithmic curation. As mediator: income gates premium tools, which gates information quality, which gates decisions. As effect modifier: decent clinical decision support buffering inexperience, or AI misinformation amplifying low health literacy. The last two remain hypotheses.
SUTVA, the stable unit treatment value assumption, breaks in two places. One person's use can rewrite a peer's feed (interference). The same daily chatbot habit is a uniquely tailored intervention for every user (multiple versions of treatment). Individual measurement is split into access, tool portfolio, intensity, purpose, and dependency; the institutional layer adds regulation, platform governance, and model properties. A worked design is given: ages 13-17, no prior depression, 12 months of chatbot use (self-report plus app logs), incident episode defined as PHQ-A ≥ 11 or a clinical diagnosis.
This is not a new RCT. The hard numbers come from the 2025 AmeriSpeak Omnibus, N = 1,163 US adults.
| Measure | Result |
| Generative AI for at least one personal purpose | 57% of US adults |
| Rarely/never use, no high-school diploma | 73%, versus 41% among BA+ |
| Daily use | 20-21% with a college degree, 8% without |
| Health-related use | 17% overall; 30% Black, 16% White, 9% Hispanic |
| Daily use by race/ethnicity | 12% Hispanic to 30% Other/multiracial (n = 112) |
The education gradient mirrors other health inequalities. A testable claim is that AI may sit on the pathway between socioeconomic status and health. Health-related use is highest among Black adults; the paper reads this as burden and access to conventional care, not simple preference. If health-facing AI helps or hurts, this group meets it first. Other/2+ leads in search (54%) and education (32%); Black adults lead in health (30%) and entertainment (31%). Race-by-purpose cells are small, and the paper flags them as suggestive. Other/2+ also bundles Asian, Native American, Pacific Islander, and multiracial respondents, so the point estimate should not be stretched.
Cited external scale: ChatGPT hit 100 million users in two months; KFF 2024 put monthly health-chatbot use at about one in six US adults; Pew reports two-thirds of US teens have used the tools; AMA reports two-thirds of US physicians used health AI in 2024. ChatGPT Health launched in early 2026, wired to records and wearables.
Existing experiments are described as low-dimensional, short-horizon, task-bound treatments. Bai et al. found LLM policy arguments as persuasive as human ones. Spitale et al. found GPT-3 disinformation hard to tell from human text. Obermeyer et al. showed a cost-trained clinical algorithm underestimating illness severity in Black patients. Those papers map mechanisms. They do not answer what chronic use does to depression and cognition, whether long-run AI advice changes care-seeking, or whether a platform-governance shift moves population mental health.
For model builders: the system is already an ambient exposure. Users do not have to open the product. Triage, claims, ranking, and poverty mapping all count. Once SUTVA fails, a per-user A/B test does not recover the population total effect.
For product teams: exposure is nudgeable, like alcohol under a tax. A safety-filter change or a new statute is a quasi-experiment. The paper asks platforms to support independent research the way drug makers support post-market surveillance, and it flags the asymmetry: pharma has mandated reporting; AI firms hold the telemetry needed to study their own products.
For public health: AI use should be measured in cohorts the way diet, activity, and screen time already are. Self-report needs calibration; digital phenotyping from phone sensors is the check. Designs named: marginal structural models, g-computation, target-trial emulation, and quasi-experiments around policy shifts.
Equity runs both ways. Missing high-quality AI is deprivation. The heaviest users, adolescents in parasocial bonds and people in crisis talking to unregulated therapy bots, may take harm first. Cost-trained clinical scores that undercount Black patients, and covert dialect bias in language models, are social structure automated and then scaled.
The authors list four. Ambient and personal blur in real use: scrolling algorithmically curated health content is both ambient curation and a personal choice. Mediator and effect-modifier roles have almost no direct tests. The algorithmic-determinant claim may swallow every ranking engine and automated decision tool, so the scope is unclear. The pollution analogy can pathologize use. Dose-response is likely non-monotone and context-dependent; use should not be classified as strictly harmful or beneficial.
A few more gaps sit in the data. AmeriSpeak is 1,163 people; race-by-purpose cells are thin. The survey describes who uses AI, and does not link use to health outcomes; the paper calls this a measurement gap. The opening "two-thirds of teens" and "one in six adults monthly" come from different KFF, Pew, and AMA vintages and do not share a definition. Dose-response, longitudinal cohorts, and instruments do not exist yet. This is an agenda paper, not an effect estimate.