State of the Prompt
How much politics
We count a conversation as political when it involves government, public policy, law, elections, or public figures. That includes people’s own dealings with government, such as filing taxes, applying for benefits, or sorting out immigration papers. By this definition, 7.8% of conversations are political, and nearly half of the people in the study (47%) had at least one political conversation during the year.
The share rises steadily with age, from 6.6% among people 18 to 24 to about 10.4% among those 55 and over. It also tends to rise with education, from 6.1% among people who did not finish high school to 9.9% among those with a postgraduate degree. That pattern largely holds when age, income, and other characteristics are taken into account, while differences by income do not.
By age
Share of each group’s conversations that are political
By education
Share of each group’s conversations that are political
Personal and public
Political conversations come in two broad kinds. Some are about the person’s own situation: a tax return, a benefit application, a visa, a ticket, or a court date. Others are about society at large: a policy debate, an election, a public figure, or a news event. The personal kind is somewhat more common.
Whose stake is at issue
Share of political conversations
The two kinds differ in subject. Personal conversations center on government operations, the law, social welfare, and the economy. Societal conversations spread more widely, with political figures, civil rights, and ideology much more prominent. Women and men raise politics at similar rates, but women’s political conversations are more often about their own affairs (57% personal, against 46% for men).
Personal conversations
Share mentioning each subject
Societal conversations
Share mentioning each subject
What it is about
Law and government operations lead by a wide margin, each appearing in more than a quarter of political conversations. Elections, immigration, and the environment, often at the center of public debate, each appear in fewer than 3%. A conversation can touch on more than one subject, so the shares add up to more than 100%.
Subjects of political conversations
Share of political conversations touching on each subject
What people do
Most political conversations are searches for information. More than a quarter are about getting something done, such as filling in a form or preparing a filing, and about a quarter seek help with a personal decision. Sharing or arguing an opinion is comparatively rare, at 7%.
What political conversations do
Share of political conversations; a conversation can do more than one thing
Over the year
The two kinds of political conversation followed different calendars. Societal conversations rose in the fall of 2025, a period that included a federal government shutdown, and declined through the spring. Personal conversations climbed in January and stayed high through April, which coincides with tax season.
Political conversations by month
Share of each month’s conversations
How people respond
In 72% of political conversations, people wrote again after ChatGPT’s first answer. Few of them challenged it. Users disputed an answer in 1.6% of political conversations, brought in an outside source in 0.4%, and asked for evidence in 0.2%. Accepting an answer was more common, though most acceptances were a brief thanks or okay.
Responses to ChatGPT’s answers
Share of political conversations with each response
When people did dispute an answer, they usually stated a different view rather than pointing to an error. Close to half (44%) gave no reason, 38% drew on their own knowledge, and 10% appealed to their values. Most disputes ended without either side giving ground.
About the data
This report draws on 3.4 million ChatGPT conversations by 65,689 adults on an opt-in US research panel, July 2025 to June 2026, through a collaboration with MFour Data Research. Conversations from before a panelist turned 18, and from panelists with no recorded age, are not included. Months are in the panelist’s local time.
A language model read the person’s side of every conversation and judged whether it was political, whose stake was at issue, and which subjects it touched. On 347 conversations coded by hand, the model’s political calls had a precision of 85% and a recall of 92%. Adjusting for those errors puts the political share near 7.2%. Subject and activity codes are less consistent than the political call itself and are best read as broad patterns.
Responses to ChatGPT were coded in two further passes that read both sides of each political conversation. 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.
Suggested citation: Ognyanova, K., Garimella, K., Aakhus, M., Jhaver, S., Lee, E. G., Li, J., & Shugars, S. (2026). Politics on ChatGPT (State of the Prompt Report No. 2). Rutgers University.