A person can be exposed to artificial intelligence without ever typing a prompt. A hospital may use an algorithm to prioritise a queue. A search engine may place an AI summary above ordinary links. A benefits system may rank a case for review. A social feed may decide which health claim receives repeated attention. Those systems can shape information, access and decisions even when the affected person does not know that a model was involved.
Researchers writing in Nature Health propose treating AI as a determinant of health and building an epidemiology of AI. Their central distinction is between personal exposure—deliberate use of an AI tool—and ambient exposure, which arrives through algorithmic curation and institutional decisions regardless of individual choice.[1]
Two kinds of exposure require different measurements
| Exposure | Example | What a study must record |
|---|---|---|
| Personal | A person asks a chatbot about symptoms | Tool, frequency, purpose, duration, response type and follow-up actionSource 1Source 2 |
| Ambient | A feed repeatedly ranks AI-generated health claims | Which system curated the content, reach, timing and who could avoid itSource 1 |
| Institutional | A hospital or insurer uses a model in a decision pathway | Model version, decision point, human override and downstream outcomeSource 1 |
| Market-wide | AI changes the supply or price of a service | Adoption timing, geographic coverage and groups with unequal accessSource 1 |
AI can occupy several positions in a causal story
The framework also warns against treating AI as one simple variable. It can be the exposure under study, such as sustained use of a mental-health chatbot. It can be a mediator, carrying the effect of income or education into the quality of information a person receives. It can be a confounder if a ranking system influences both the content people see and the health behaviour being measured. Or it can modify an existing relationship—for example, changing how clinician experience affects diagnostic accuracy.[1]
Why measurement is already urgent
| Measure | Value | Context |
|---|---|---|
| Health-chatbot use | 32% of US adults in the KFF poll | Respondents said they had used AI for physical or mental health information in the previous year.Source 2 |
| Personal data sharing | 41% of AI-for-health users | They reported uploading medical information such as test results or clinicians’ notes.Source 2 |
| Access pressure | 19% cited inability to afford a professional as a major reason | The share was higher among younger and lower-income users.Source 2 |
| General personal AI use | 57% in a nationally representative Brookings/NORC survey | Use varied strongly with education, age and income.Source 3 |
A randomised trial cannot capture every population effect
Short experiments are valuable for testing one interface or message, but they can miss chronic exposure, spillovers and unequal adoption. If one person’s AI-generated post changes what thousands of others see, the usual assumption that each participant’s treatment is isolated has already failed. If premium tools reach affluent groups first, a study of average benefit may conceal a widening gap.
The proposed epidemiology would therefore need longitudinal data, version histories, exposure logs and linkage to health and social outcomes. It would also need governance that permits scrutiny without creating a new surveillance system. Measuring ambient exposure is difficult precisely because it lives inside private ranking systems and institutional workflows.
Sources & further reading
- The epidemiology of artificial intelligence Nature Health · 15 Sep 2026 · Accessed 16 Oct 2026
- Poll: 1 in 3 Adults Are Turning to AI Chatbots for Health Information KFF · 25 Mar 2026 · Accessed 16 Oct 2026
- How are Americans using AI? Evidence from a nationwide survey Brookings Institution · 1 Dec 2025 · Accessed 16 Oct 2026