State of the Prompt
Why it matters
While Americans increasingly get information from conversational AI, these exchanges tend to remain private. Unlike news broadcasts or social media feeds, the content of AI chats is largely unavailable to journalists, regulators, and researchers. This project will examine how Americans use conversational AI for information and advice, from following the news to navigating health, work, and family life.
We look at who uses AI for what and how that differs by age, education, income, and region. We also trace how these patterns shifted over a year in which conversational AI became part of daily routines for many Americans.
A year of conversations
Through a collaboration with MFour Data Research, the project draws on ChatGPT conversations from 73,852 members of a US consumer research panel. Panelists agreed to share their conversations through the panel’s app. Between July 2025 and June 2026, they had 3.9 million conversations with ChatGPT, made up of 39.8 million messages. For most panelists, the conversations are linked to a self-reported profile that includes age, gender, education, income, race and ethnicity, and state of residence.
The panel is not a random sample of the country. It includes more women and more people with lower incomes than the US population as a whole. For that reason, our analyses focus on comparisons between groups rather than on national estimates.
Who is in the figures
Everything on this site describes adults. The figures cover 65,689 panelists with a demographic profile, and count only the conversations they had at 18 or older: 3.4 million conversations in all. Conversations by teenagers, and by panelists whose age was not recorded, are left out. Age groups use each panelist’s age in July 2026, when the demographic profile was taken, so a conversation from earlier in the year can count in the next age group up. Times of day are the panelist’s local time, based on the state they live in.
Reading at scale
No team could read millions of conversations, so we rely on language models to sort them. The models classify each conversation by subject, purpose, and political content, and a topic model adds a finer map of about 100 recurring themes. Where possible, we check the models’ judgments against conversations coded by members of the team.
What stays private
These exchanges are often deeply personal. They include health questions, family conflicts, and money worries, and some contain names or other identifying details. The full conversations are stored on secure university systems and are available only to the research team, under a license that permits research use at Rutgers.
Everything published on this site is an aggregate. Every figure describes a group of people, never an individual, and no conversation is ever shown or quoted. Figures that rest on too few conversations, or on a handful of people, are withheld.
