How people actually use AI chatbots

The picture most people have of AI chatbots comes from work: drafting emails, writing code, summarizing meetings. The usage data says something else. When OpenAI's economists classified a sample of about 1.1 million ChatGPT conversations, most messages had nothing to do with a job, and the share kept growing. People ask chatbots what to do, how something works, what a word means and whether a rash is serious. Coding, the use that dominates the conversation about AI, is a small slice of ChatGPT and a large slice of Claude.

We put the vendors' own usage studies (OpenAI, Anthropic, Microsoft, Perplexity) next to independent surveys (Pew, the Real-Time Population Survey, Gallup, the Reuters Institute, Menlo Ventures) and traffic data from Similarweb. Each vendor describes only its own users and sorts them with its own classifier, and each survey asks a slightly different question, so every figure says whose numbers it shows and what it counts. The methodology has the details.

Mostly not work#

The largest study of chatbot use is OpenAI's own. In "How People Use ChatGPT", a September 2025 NBER working paper, OpenAI's economic research team and Harvard's David Deming classified one message from each of about 1.1 million sampled conversations sent between May 2024 and June 2025. No human read the messages. The sample covers the consumer Free, Plus and Pro plans only: Business, Enterprise and Edu accounts, users who opted out of training, users under 18 and logged-out users aren't in it.

Over that year, messages that weren't about work grew much faster than messages that were. In June 2024 ChatGPT received about 451 million messages a day, 53% of them non-work. A year later it received 2.6 billion, and 73% were non-work. The authors write that the shift happened mostly within cohorts of users, not through an influx of new ones: people who already used ChatGPT brought more of the rest of their lives to it.

The work share still varies a lot by who's typing. Among users who shared employment data, 57% of messages from people in computer-related jobs were work-related, against 40% for people in non-professional jobs. Users with a graduate degree sent 48% work messages, users without a bachelor's 37%. And the sample is young: about 46% of messages from users who report an age come from 18 to 25-year-olds.

What people ask for#

OpenAI groups conversations into seven topics. Three of them, practical guidance, seeking information and writing, account for about 77% of conversations. Practical guidance (tutoring, how-to advice and the like) held at about 29% of usage between July 2024 and July 2025. Writing fell from 36% to 24%, while seeking information, which looks most like web search, rose from 14% to 24%.

Writing is about 40% of work-related messages, and most writing is editing: about two-thirds of all writing messages ask ChatGPT to change text the user provided (editing, critiquing, translating, summarizing) rather than produce something new. Tutoring and teaching are 10.2% of all messages. The paper also sorts messages by intent: by late June 2025, 51.6% were "asking" (looking for information or advice), 34.6% "doing" (producing an output) and 13.8% "expressing". In July 2024 asking and doing had been roughly even.

Surveys of the public find the same order. In Pew's February 2026 survey of 5,119 US adults, searching for information was the most common use of chatbots, ahead of work tasks.

The Reuters Institute's six-country survey (Argentina, Denmark, France, Japan, the UK and the US) reached the same conclusion for mid-2025: weekly use of generative AI to get information rose from 11% to 24% of people in a year, and "information-seeking is now the lead use-case across countries", ahead of creating media.

Health is the clearest example of everyday advice. OpenAI said in January 2026 that health is a little over 5% of ChatGPT messages and that more than 40 million people a day ask it health questions, as reported by Axios. Microsoft's analysis of 37.5 million consumer Copilot conversations, "It's About Time", found that on phones "health is the dominant topic", in every hour of the day and every month it observed, while on desktop "Work and Career" overtook "Technology" as the top topic during business hours. In a follow-up on health conversations, Microsoft found one in seven questions about symptoms and conditions were asked on behalf of someone else.

Coding concentrates on Claude#

On ChatGPT, computer programming is 4.2% of messages. The paper's broader "technical help" topic, which also covers math and data analysis, fell from 12% of usage in July 2024 to about 5% a year later. The authors suggest one reason: programming with language models has grown fast elsewhere, through the API, code editors and agents such as Codex.

Anthropic's numbers show where some of it went. In its Economic Index, which classifies a sample of Claude conversations each release, coding was 36% of Claude.ai conversations in August 2025, and computer and mathematical tasks were 44% of the traffic from businesses using Anthropic's API.

Claude.ai's users skew toward developers, so this says as much about who uses Claude as about what chatbots are for. Even so, Anthropic's coding share has stayed high: computer and mathematical tasks peaked at 40% of Claude.ai conversations in March 2025 and were 34% in November 2025. Since then Anthropic says coding has been moving from Claude.ai to the API and Claude Code, and its June 2026 report moved to a new pipeline that reports chat and Cowork together and Claude Code separately, so its later numbers aren't comparable. For what coding agents do once they leave the chat window, see the state of AI coding agents.

Claude is getting more personal too#

Anthropic sorts Claude.ai conversations into work, coursework and personal use. Between November 2025 and February 2026, the personal share rose from 35% to 42% and coursework fell from 19% to 12%, while work stayed at about 45%. Anthropic also found that people who signed up a year earlier devote 38% of their conversations to personal use, against 44% for the newest users, so new users arrive with more personal questions.

The Economic Index also tracks how people work with Claude. "Automation" means the user hands over a task (directive or feedback-loop conversations); "augmentation" means the user works through it with Claude (iterating, learning, validating). Augmentation led from the first report. In August 2025, automation passed it for the first time, but Anthropic notes that part of that jump came from switching classifier models: re-run with the older one, automation was 45%, not 49%. By November it had fallen back.

Perplexity's study of its Comet browser agent points the same way from a different product: 55% of agent queries between July and October 2025 were personal, 30% professional and 16% educational.

Feelings are a small share of the logs#

A popular claim is that therapy and companionship are now the main use of chatbots. The logs don't show it. On ChatGPT, relationships and personal reflection are 1.9% of messages and games and role play 0.4%. On Claude.ai, Anthropic's study of support and companionship found 2.9% of conversations were "affective" and companionship and roleplay together under 0.5%.

Surveys measure something else: whether a person has ever done it, not what share of their chats it is. By that measure it's common. Pew found 10% of US adults use chatbots for emotional support or advice, and 20% of adults under 30. An Elon University survey run by YouGov in May 2026 found 27% of US adults who use the internet have significant emotional or social interactions with chatbots. Both can be true: many people have the occasional personal conversation, and it's a small fraction of the total.

Half of Americans, a quarter daily#

Pew has asked Americans about ChatGPT since 2023. The share who say they use it rose from 18% in 2023 to 34% in 2025. In 2026 the question also asked about other chatbots, and 44% said they use ChatGPT; the earlier waves were asked only of people who had heard of ChatGPT, so the last step isn't a like-for-like comparison. Among adults 65 and older the share nearly doubled in the last wave, though the new wording affects that step too.

Counting any chatbot, 49% of US adults use one, up from 33% in 2024, and the gap between men and women has closed (50% and 47%). The Real-Time Population Survey by Alexander Bick, Adam Blandin and David Deming, which asks US adults aged 18 to 64 about any generative AI every three months, shows the same rise and where it comes from. Use for any purpose went from 44.6% in August 2024 to 60.8% in August 2026. Use outside work rose in every one of the nine waves, from 36.0% to 55.1%.

Daily use is where the numbers stay small. Every source we have puts daily users at a quarter of its population or fewer, except teenagers.

Reach is growing more slowly than intensity. In Menlo Ventures' surveys with Morning Consult, the share of US adults who used AI in the past six months rose only from 61% to 64% between April 2025 and July 2026, while daily use rose from 19% to 25%. Among non-users, Pew found 83% say they aren't interested, 79% cite privacy and 76% don't trust the accuracy; 67% say they're not likely to start in the next year. If privacy is the worry, what ChatGPT does with your data is the place to start.

At work, rising but unevenly#

Use at work is growing too. Gallup's quarterly panel of US employees found 52% use AI in their role at least a few times a year in May 2026, up from 21% in 2023, and 15% use it daily, up from 4%.

The RPS puts work use at 40.5% of employed adults in August 2026, ranging from 68.8% in computer and math jobs to 13.5% in building and grounds work, and estimates that 5.3% of all work hours were spent using generative AI. The two surveys disagree on the latest direction: Gallup kept rising through May 2026, and the RPS dropped in August. Gallup's next wave hadn't been published when we wrote this.

ChatGPT is bigger, its share is smaller#

OpenAI's user counts have kept climbing. ChatGPT had 100 million weekly users when Sam Altman announced it in November 2023, 800 million in October 2025 and, OpenAI said in its 31 August 2026 post on ads, more than 1 billion.

Everyone else grew faster on the web. Using the monthly visits Similarweb published for 2025, we computed each site's share of the combined visits to seven chatbot domains. Between January and December, ChatGPT's visits rose 43%, Gemini's 548% and Claude's 125%. DeepSeek's spike in February 2025 faded within months.

The shift continued into 2026. Similarweb wrote in July 2026 that ChatGPT's share of generative AI website visits fell from about 76% in June 2025 to roughly 53% by May 2026, while Gemini rose from under 9% to around 27–28% and Claude "moved from barely 2% to close to 9%". Since then ChatGPT has recovered a little. Similarweb's September 2026 numbers, reported by OfficeChai, show ChatGPT at 56.4% and Gemini at 25.0%, Gemini's fourth monthly decline in a row.

Surveys show the same fragmentation from the user's side. In Pew's February 2026 survey, 44% of US adults had used ChatGPT, 24% Gemini, 17% Copilot, 14% Meta AI, 8% Grok and 6% Claude. People who use AI tend to use several: in Menlo's July 2026 survey, AI users used 3.0 general assistants on average, up from 2.2.

What this means if you build with AI#

1. Design for advice and information, not only for tasks. Practical guidance and information-seeking are over half of ChatGPT use, and "asking" outgrew "doing". If your product sits inside or beside a chatbot, the common request is "help me decide" or "explain this", so make answers easy to check and to act on.
2. Expect people to bring their own text. Two-thirds of writing requests modify text the user supplied. Features that take a pasted draft, a file or a link and keep the person's voice match how people already write with chatbots.
3. Build for weekly users, not daily ones. At most a quarter of adults use a chatbot daily. Most of your users will come back after days away, so make the first answer good without setup and keep context they'd otherwise have to repeat.
4. Go where your users' kind of work happens. Developers are concentrated on Claude, the API and coding agents; mass-market questions go to ChatGPT and, increasingly, Gemini. A developer tool belongs in Claude Code, Cursor and MCP directories; a consumer tool needs to work where most of the 1 billion weekly users are.
5. Don't bet on one assistant. People use three on average and the web share moves by tens of points in a year. Build on open interfaces (MCP, skills, plain links and files) so the same tool works in ChatGPT, Claude and Gemini.
6. Treat health, money and feelings as core use, with care. Health is the top Copilot topic on phones and 20% of US adults have asked a chatbot for medical advice. Say what your product can't do, point to a person when it matters, and expect questions asked on someone else's behalf.
7. Answer the privacy question up front. Privacy (79%) and accuracy (76%) are among the top reasons non-users give in Pew's survey. Say what you store, for how long, and whether it trains a model.
8. Measure your own users. Every vendor number here describes that vendor's users through its own classifier. Your mix of work and personal use, and your share of coding, will differ; log categories (privately) before you plan around anyone else's.

When a chat produces something worth keeping, such as a page, a deck or a small app, host0 hosts it at a link that stays private until you share it.

Methodology#

Sources. Vendor usage studies: OpenAI and Harvard's NBER working paper w34255 (about 1.1 million sampled consumer conversations, May 2024 to June 2025, classified by a model, no human reading); Anthropic's Economic Index (six releases from February 2025 to June 2026) and its June 2025 study of support and companionship; Microsoft AI's Copilot usage report (37.5 million consumer conversations, January to September 2025) and its health report; Perplexity and Harvard's study of the Comet agent. Surveys: Pew Research Center (American Trends Panel, February 2026, n = 5,119; teens, Ipsos KnowledgePanel, September–October 2025, n = 1,458); the Real-Time Population Survey (US adults 18–64, nine waves, August 2024 to August 2026); Gallup's panel of US employees; the Reuters Institute's six-country survey (YouGov, June–July 2025); Menlo Ventures with Morning Consult (July 2026, n = 5,067); Elon University with YouGov (May 2026). Traffic: Similarweb's published visit estimates and share snapshots.

What we computed. The Claude.ai collaboration and use-case shares come from the global rows of each Economic Index release on Hugging Face: automation is directive plus feedback-loop conversations, augmentation is task iteration, learning and validation. The first release didn't sum to 100%, so we normalized it. The RPS series come from the authors' workbook rather than FRED, because FRED's "adoption rate outside of work" series matches the workbook's at-home use by people without a college degree, not overall non-work use. The 2025 web shares are each domain's share of Similarweb's monthly visits to seven chatbot domains.

Limits. Vendor studies describe their own users only: ChatGPT's paper covers consumer plans, not Business or Enterprise, and Claude.ai's users skew toward developers. OpenAI counts messages and Anthropic conversations, and each uses its own taxonomy, so cross-vendor comparisons show the size of a gap, not an exact difference. Survey questions differ (ever used, used in the past six months, used last week), as do populations (all adults, ages 18–64, employees, teens). Similarweb measures website visits, not users, and leaves out apps and API traffic; its monthly shares after June 2025 come from its posts on social media as reported by OfficeChai, FourWeekMBA and others, and outlets differ by a point or two for the same month. We re-read the quoted lines in the primary documents on 24 October 2026. The data pack, with every row, its source and tier, the scripts and the conflicts we resolved, is in our research notes for this post.

Open questions#

• What do people ask Gemini? Google publishes user counts for the Gemini app but no breakdown of topics or work and personal use.
• Has ChatGPT kept moving away from work? OpenAI's public work and topic series ends in mid-2025; its January 2026 health figures are the only newer topic share.
• Is work use falling? The RPS dropped from 45.2% to 40.5% in one wave while Gallup still rose. The next waves will show whether it was noise.
• How much happens in apps? Every public share series measures website visits. No independent source publishes app usage shares across chatbots.
• What's the work share across all plans? ChatGPT's paper excludes Business and Enterprise accounts, so the overall split of work and personal use on ChatGPT is unknown.