State of the Prompt
How much is about health
Report 1 counts conversations whose main subject is health and medicine: one in ten. This report asks a broader question of every conversation with readable text: is it about health at all, including emotional support, diet and fitness, and finding care? Two model checks have to agree before a conversation counts.
The two measures are not interchangeable. Report 1 reads the main subject of a conversation; this page reads whether health comes into it at all. Both are shown on this site, each with its own definition.
What kinds of question
Each health conversation was sorted into one of twelve kinds, message by message, with the conversation taking the majority. Emotional support is the largest kind. The classic case, a symptom right now or a diagnosed condition, is about a quarter. Switch between everyone, women and men to see how the mix changes.
Kinds of health conversation
Share of health conversations, %
The questions change with age. Emotional support falls away after 55, while general information and the practical questions about finding care, paperwork and insurance rise. Studying is a young person’s use. Each panel below is one kind of question, read left to right from the youngest adults to the oldest.
Kinds of question by age
Share of each age group’s health conversations, %. Each panel has its own scale, with every value labelled.
Who asks
Women’s conversations are about health almost twice as often as men’s, and the gap holds after adjusting for age, income, education, employment, ethnicity, region and children. By age, health use peaks at 25 to 34. Income makes no difference to how much people use ChatGPT for health, and education very little; what they do shape is the kind of question, two sections on. Choose a kind of group.
Share of each group’s conversations that are about health
Health as a share of all the group’s readable conversations, %
Men and women ask about different things. Women’s questions are clinical and emotional; men’s are informational and about fitness. Adjusted for everything else, a woman’s health conversation is 1.4 times as likely to be about her own symptoms or a condition, and about two thirds as likely to be about diet and fitness or general information. Women also come back more: 20 health conversations per health user against 13 for men.
Kinds of question, women and men
Share of each group’s health conversations, %
WomenMen
Who the question is about
Half of health conversations are about the user. Three in ten are impersonal: how insulin works, what a test measures. One in seven is about another adult, a partner, parent or friend, and often it is the emotional side of that relationship rather than that person’s care. One in fourteen is about the user’s child or dependent.
Who the health question is about
Share of health conversations
Health care for children in these households is being routed through ChatGPT, mostly by mothers. Women with children at home ask about a child in 12% of their health conversations, men with children at home in under 5%. The gap holds in every age band, every income and education level, every employment status, every region and every ethnicity; the ratio never drops below 1. Choose a view.
Health conversations about a child or dependent
Share of the group’s health conversations, %
Income and education
Income does not change how much people use ChatGPT for health, but it changes what they bring to it. Emotional support falls steadily as income rises, and diet and fitness questions rise in its mirror image. Symptom and condition questions stay within about two points across all six income bands: income reshapes the emotional and lifestyle ends of health use and leaves the clinical middle alone. Education shows the same gradient in emotional support.
Where the gradient is
Share of the group’s health conversations, %
Life stage shows in the questions
Employment, relationship status and children at home each leave a mark on what people ask.
How people ask
People describe the symptom; they rarely ask whether to see a doctor. “My throat hurts” and “it’s been swollen for days” outnumber “should I see a doctor” three to one. People hand ChatGPT the symptom and let it decide what to do with it. A tool that only triggers on “should I see a doctor” would miss most of what is going on.
About the data
Every conversation with at least one message and reply in text went through the same two steps, run by an open language model (Gemma 3, 27B) on Rutgers’ cluster, so no conversation text left it. First, a yes-or-no check: is this about health? Second, a classifier of twelve kinds of health use, applied message by message, with the conversation taking the majority. A conversation counts as health only when both agree. Image-only conversations, about 200,000, cannot be read and are left out.
This page covers adults only: conversations by panelists 18 or older. Figures by age, gender, income, education, employment and parenthood use each person’s self-reported profile. Shares are of conversations unless noted, so people who use ChatGPT heavily count for more. The panel is not a random sample of the country, so the figures are best read as comparisons between groups.
The definition here differs from Report 1, which uses the main subject assigned to each conversation from a list of 28. Both are published on this site with their definitions stated. “Health” here means two model checks agree; a human check of the definition is in progress.
Suggested citation: Garimella, K., Ognyanova, K., Aakhus, M., Jhaver, S., Lee, E. G., Li, J., & Shugars, S. (2026). How Americans talk to AI about health and emotional support (State of the Prompt Report No. 3). Rutgers University.