How Are People Using AI in 2026?

Key Takeaways

• People use AI most for information, practical guidance, writing, learning, and work tasks.
• Workplace use is shifting from drafting and search toward delegated, agentic execution.
• Adoption is broadening, but age, income, occupation, geography, and skill still shape use.

What the Best Research Says About How People Are Using AI

On August 6, 2026, OpenAI reported that more than 1 billion people were putting ChatGPT to use, and its country-level data showed a measurable shift from asking questions toward getting tasks completed. That finding captures how people are using AI after almost four years of mass-market generative artificial intelligence (AI): people still ask for information, explanations, and advice, but they increasingly ask systems to write, analyze, create, research, program, edit, and act on their behalf.

The evidence base has become much stronger than it was during the early ChatGPT boom. OpenAI has analyzed millions of privacy-protected conversations. Anthropic has classified large samples of Claude conversations and business application programming interface traffic. Microsoft Research has examined hundreds of thousands of Copilot conversations and conducted workplace experiments. Pew Research Center, Gallup, Stanford University’s Institute for Human-Centered Artificial Intelligence, Common Sense Media, the American Medical Association, and university research groups have added surveys and controlled studies covering consumers, employees, students, teachers, physicians, and technical professionals.

These sources do not measure exactly the same thing. Platform telemetry shows what users actually request from a particular service but cannot represent people who do not use that service. Surveys can reach nonusers but depend on memory and self-reporting. Controlled experiments can estimate causal effects on performance but usually cover narrowly defined tasks. Research on how people are using AI needs all three forms of evidence.

Stanford’s 2026 AI Index estimates that generative AI reached 53% population adoption within three years, faster than the historical diffusion of either personal computers or the internet under the report’s comparison methodology. The same research estimates U.S. consumer surplus from generative AI at $172 billion annually by early 2026, up from $112 billion a year earlier. The underlying Stanford Digital Economy Lab research found that the median value users placed on monthly access rose substantially between the study periods. These figures indicate that consumer value is being created well beyond paid subscriptions because many widely used services remain free or inexpensive.

OpenAI’s large 2025 consumer study offers one of the clearest classifications of actual behavior. Based on a privacy-preserving analysis of 1.5 million conversations, roughly 70% of consumer ChatGPT use was unrelated to work. Practical guidance, information seeking, and writing together accounted for approximately three-quarters of conversations. About 49% of messages were classified as “Asking,” 40% as “Doing,” and 11% as “Expressing.”

That result corrects a common misconception. Generative AI adoption cannot be understood primarily as workplace automation. Consumers use these systems as search substitutes, explainers, tutors, writing assistants, planners, creative tools, health information services, troubleshooting aids, and decision-support systems. Work matters enormously, but personal use accounts for much of the observed activity.

New Space Economy’s examination of AI in Americans’ online experience adds another measurement issue. AI use increasingly occurs inside search engines, browsers, productivity software, shopping platforms, phones, and other services. People can now consume an AI-generated answer without deliberately opening a chatbot. The boundary between choosing to use AI and using software that contains AI has become difficult to measure.

Everyday AI Use Is Dominated by Information, Advice, Planning, and Creation

Pew Research Center’s February 2026 survey of 5,119 U.S. adults offers an unusually detailed picture of stated consumer behavior. Forty-two percent said they had used AI chatbots to search for information. Twenty-five percent used them for entertainment, 24% for creating or editing images and videos, 20% for medical advice, 20% for diet and fitness information, and 13% for news.

Those categories overlap with OpenAI’s observed conversation patterns. The largest consumer uses tend to begin with a need rather than a technology: explain something, recommend an approach, rewrite a document, compare choices, troubleshoot a problem, plan an activity, summarize material, or produce something new.

Information search deserves particular attention because it changes one of the internet’s most common behaviors. Traditional web search generally returns links that users evaluate. Generative systems can instead synthesize information into an answer and then continue the interaction through follow-up questions. The user can request a shorter explanation, ask for a comparison, supply additional facts, or challenge the response. Search becomes conversational.

A Pew Research Center tracked-browsing study from March 2025 found that 58% of its 900 U.S. participants encountered at least one search query that generated an AI summary. Only 13% directly visited a generative AI tool website during that month. That difference shows how exposure can expand much faster than intentional chatbot visitation.

Planning and practical guidance also account for considerable activity. People ask AI systems about travel, cooking, home maintenance, purchasing decisions, schedules, hobbies, interpersonal communication, job applications, financial concepts, and administrative problems. Anthropic’s Canadian research, based on February 2026 Claude usage data, found substantial personal use across Canadian provinces, with examples including product research, health information, recipes, and home repairs.

Creative use is becoming more visual. OpenAI reported in August 2026 that multimedia generation, analysis, and retrieval had grown to 7.8% of classified ChatGPT messages globally during Q2 2026. In Brazil and Colombia, multimedia represented more than 10% of messages. Images are becoming part of ordinary AI use rather than a separate specialty application.

Writing remains one of the strongest recurring behaviors because language appears in so many areas of life. Users draft correspondence, correct grammar, alter tone, translate passages, summarize documents, brainstorm ideas, prepare applications, and convert rough notes into structured prose. OpenAI’s research found that editing or transforming text represented a large share of writing use rather than users simply requesting complete original documents.

Usage also deepens with experience. OpenAI’s June 2026 cohort analysis found that six months after account creation, sampled users were sending about 50% more messages per day than when they joined and had doubled the number of distinct tasks they had attempted. Anthropic found a related pattern in its March 2026 Economic Index: users with six months or more of Claude experience had a 10% higher measured conversation success rate than newer users after researchers accounted for several observable differences.

These findings suggest that adoption has two dimensions. More people begin using AI, and established users discover additional jobs for it. The latter effect matters because a service can gain economic significance even after headline user growth slows.

How People Are Using AI at Work Is Shifting From Assistance to Execution

Workplace use began with low-friction tasks that already happened on computers: drafting emails, summarizing meetings, rewriting documents, finding information, creating presentations, analyzing text, writing formulas, and helping with code. New Space Economy’s analysis of day-to-day generative AI work describes this pattern as draft, check, rewrite, and route, with employees using AI as a working surface before human review.

Microsoft Research reached a similar finding after analyzing 200,000 anonymized Bing Copilot conversations. Users most commonly sought assistance with gathering information and writing. The system itself most frequently performed activities involving information provision, writing, teaching, and advising. Occupations centered on information processing and communication consequently showed higher measured AI applicability.

OpenAI’s 2025 consumer research estimated that about 30% of ChatGPT consumer-plan use was work-related. Writing dominated the professional subset. By 2026 the work pattern was becoming more specialized, with content creation, documentation, information retrieval, analysis, and software-related activity extending into more workplace functions.

Enterprise usage is moving further. OpenAI’s August 12, 2026 enterprise research indicates that organizations at the upper end of AI intensity increasingly use agents, systems that can perform multi-step work with access to tools and organizational context. In June 2026, Codex accounted for 64% of combined Codex and ChatGPT output tokens among enterprise customers. Output-token volume should not be interpreted as an employee adoption rate because agentic tasks naturally generate more output, but the change documents a shift in workload composition.

The functional spread is striking. From February through the measurement period, weekly active enterprise Codex users increased 108-fold in legal, 41-fold in sales, 41-fold in recruiting, and 26-fold in marketing. Engineering increased fivefold from a much more established base. The underlying tasks differ by department: researching and drafting legal material, preparing customer information, processing recruiting work, generating marketing assets, or operating software-related tools.

OpenAI also found a pronounced gap among organizations. Firms in the top 10% of measured AI usage generated 8.3 times as many output tokens per active user as firms near the middle of the distribution in June 2026, compared with 2.6 times in January. Twenty-one percent of active users at the high-use organizations used Plugins each week, versus 9% at typical firms. OpenAI’s Enterprise Signals dataset provides the underlying measurements and cautions that token volume is an imperfect proxy for business value.

Stanford’s 2026 AI Index puts this behavior into a wider business frame. Its economy chapter reports that 88% of surveyed organizations used AI in at least one business function in 2025. Generative AI was used in at least one business function at 70% of surveyed organizations, although deployment of AI agents remained in the single digits across nearly all measured business functions.

Access does not equal deep integration. Many employees still use AI intermittently for isolated tasks. A smaller group is redesigning workflows around models, connected data, tools, automation, and human approval. This difference increasingly separates casual organizational adoption from operational adoption.

Coding and Technical Work Show Both the Power and Limits of AI Productivity Claims

Software development remains one of the most concentrated areas of generative AI use. Anthropic’s January 2026 Economic Index found that computer and mathematical tasks represented about 34% of Claude.ai conversations and 46% of first-party application programming interface traffic in November 2025. Software error correction and other coding-related tasks accounted for a substantial part of this activity.

Coding is suited to AI because much of the work exists in machine-readable form and can be tested. Models can generate functions, explain unfamiliar repositories, write tests, refactor code, diagnose errors, create scripts, translate between languages, prepare documentation, and operate development tools. Agentic systems extend that model by reading files, editing repositories, executing tests, and repeating attempts.

The human role is changing accordingly. Technical workers increasingly spend time describing objectives, dividing work, reviewing generated code, testing outputs, deciding whether an agent took a sensible approach, and resolving failures that automatic tests cannot detect. METR’s research program on frontier AI systems has documented the growing capability of coding agents alongside the continuing difficulty of translating technical capability into a single reliable productivity estimate.

Productivity evidence remains mixed enough to require restraint. A widely cited workplace study covering 5,179 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by about 14% on average. Gains reached 34% among novice and lower-performing workers, with little measured benefit for the most experienced workers. The study was later published in the Quarterly Journal of Economics after appearing as an NBER working paper.

Microsoft’s six-month randomized field experiment involving 6,000 knowledge workers found a different form of change. Workers who actually used the generative AI tool spent about three fewer hours per week on email and appeared to complete documents faster, but meeting time did not change significantly. The Microsoft Research field experiment also distinguishes between the effect of receiving access to the technology and the larger effect among workers who actually used it.

Software experiments have produced less uniform results. METR’s early-2025 randomized study of experienced open-source developers found that participants took 19% longer when using the available AI tools, even though the developers believed AI had made them faster. Later tools became more capable, but researcher selection problems and changing user behavior make comparisons across model generations difficult.

A May 2026 METR survey of 349 technical workers produced much larger self-reported gains. Respondents reported substantial increases in the value and speed of their work. METR explicitly warned that self-reported counterfactual productivity is difficult to estimate and may exceed experimentally measured gains.

The evidence supports a narrower interpretation than either enthusiasts or skeptics often present. AI can produce large gains on suitable tasks. Effects vary by occupation, expertise, model quality, task structure, and measurement method. Faster output does not automatically produce proportionately higher economic value.

Students and Educators Have Made AI Part of Learning

Education may contain some of the highest sustained AI usage rates of any large social institution outside technology-intensive workplaces. Stanford’s 2026 AI Index education findings report that more than 80% of U.S. high school and college students use AI for school-related work. Research, writing assistance, brainstorming, and other academic tasks have become regular applications.

The United Kingdom provides even stronger adoption evidence. The Higher Education Policy Institute’s Student Generative AI Survey 2026 of 1,054 full-time undergraduates found that 95% had used AI in at least one way and 94% used generative AI to help with assessed work. Direct inclusion of AI-generated text in assessed work rose to 12%, compared with 8% in 2025 and 3% in 2024.

Students use AI for more than answer generation. Common applications include explaining difficult concepts, summarizing articles, proposing research ideas, brainstorming, editing, translating, creating study materials, analyzing information, checking work, and receiving immediate feedback.

Usage does not automatically produce better learning. Common Sense Media’s nationally representative August 2026 survey of 1,017 U.S. teenagers found that 70% used AI for schoolwork. Among those users, 63% reported obtaining schoolwork answers from AI in some form. At the same time, 77% used AI for schoolwork assistance that did not involve simply requesting an answer, including brainstorming and checking work.

The distinction matters. An AI tutor can explain an equation repeatedly, adapt vocabulary, generate practice questions, compare two interpretations of a text, or identify weaknesses in a draft. The same system can supply an answer before the student attempts the problem. Research is increasingly concerned with the sequence of use rather than a simple use-versus-nonuse distinction.

The teenagers surveyed by Common Sense Media recognized part of that tradeoff themselves. Thirty-eight percent of AI-using teens said using it caused them to generate fewer of their own ideas, and 39% said they felt they were missing learning when AI completed assignments. Only 27% said a teacher had discussed what AI is and how it works.

Educators are adopting the same tools. Anthropic analyzed about 74,000 Claude conversations associated with higher-education professionals during May and June 2025. Its educator usage study found that faculty used AI to develop course material, draft grant proposals, advise students, manage administrative tasks, create interactive teaching tools, and prepare grading materials. Routine administrative work tended to contain more automation, whereas teaching and advising involved more human-AI iteration.

Education research is consequently moving away from the question of whether students will use AI. In many institutions that question has already been answered. The harder problems involve assessment design, disclosure, verification, subject mastery, independent reasoning, faculty practice, and deciding when assistance improves learning rather than replacing it.

Health, Medical Advice, and Emotional Support Have Become Significant Uses

Health-related AI use has moved beyond speculative demonstrations. Pew’s February 2026 U.S. survey found that 20% of adults said they had used AI chatbots for medical advice and another 20% for diet or fitness information. Those figures measure self-reported experience rather than clinical appropriateness, but they place health among the common consumer applications of conversational AI.

A 2026 Nature Health study examined more than 500,000 de-identified Microsoft Copilot conversations involving health during January 2026. Nearly one in five contained personal symptom assessment or discussion of a medical condition. One in seven personal health queries concerning symptoms or conditions related to someone other than the user, indicating use by caregivers as well. Personal queries increased during evenings and nighttime, when access to conventional care can be more limited.

Consumer health use includes understanding symptoms, interpreting terminology, learning about treatments, preparing for medical appointments, generating questions for clinicians, understanding medications, discussing nutrition, and navigating health systems. These uses sit on a spectrum from education to clinical decision support, and the safety requirements change sharply along that spectrum.

Physicians are also adopting AI at high rates. The American Medical Association’s 2026 survey of 1,692 U.S. physicians found that 81% reported using AI professionally, compared with 38% in 2023. Thirty-nine percent used it to summarize medical research or standards of care, 30% for discharge instructions or care plans, 28% for billing or visit documentation, 28% for chart summaries, and 19% for draft replies to patient portal messages.

Clinical adoption shows why AI use cannot be measured only by chatbot conversations. Some systems operate inside electronic health records, transcription tools, imaging systems, administrative software, and clinical decision-support products. A physician may use AI without opening a general-purpose chatbot at all.

Personal and social use deserves separate attention. Pew found that 10% of U.S. adults had used chatbots for emotional support or advice and 4% for companionship. Younger adults reported higher emotional-support use than older groups, making age one of the strongest demographic variables in this category.

Teen behavior is more pronounced. Common Sense Media’s 2025 AI companion study found that 72% of surveyed teenagers had used AI companion products at least once and 52% used them regularly. One-third had chosen an AI companion instead of a person for a serious conversation, and one-quarter had shared personal information with such systems.

These figures do not establish that chatbot companionship improves or harms well-being in every case. They establish that emotional interaction is a real use category rather than an edge case. Research now has to distinguish occasional advice, entertainment, role-playing, emotional disclosure, persistent companionship, and dependence because those behaviors carry different implications.

AI Adoption Is Broadening but Remains Uneven

The stereotypical early generative AI user was young, highly educated, technically skilled, and located in a wealthy country. That profile still describes disproportionately heavy use in some datasets, but adoption is spreading beyond it.

Pew found in 2025 that 34% of U.S. adults had used ChatGPT, roughly double the 2023 share. Work use showed substantial educational differences: 45% of employed adults with postgraduate degrees reported workplace ChatGPT use, compared with 17% among workers with a high school education or less. Younger employees also reported higher use. Pew’s February 2026 data subsequently placed reported ChatGPT use at 44% of U.S. adults, although the organization noted a change in question wording between survey years.

OpenAI’s actual usage data suggests those gaps are changing. Users over 35 increased their share of ChatGPT messages in nearly every measured country during the year ending in Q2 2026. In France and Czechia, the share of messages from users aged 35 and older rose by more than 10 percentage points compared with Q2 2025. Countries in Latin America, Africa, and Oceania also moved upward in per-capita usage rankings.

Anthropic sees strong income effects. Its Economic Index research finds Claude usage per capita strongly associated with national income, with higher-income countries generally using the product more intensively. Usage also changes as adoption matures. Less-developed markets can show proportionately more education and specialized professional use, whereas richer and more mature markets exhibit greater personal-use diversity.

This resembles diffusion patterns seen with earlier general-purpose technologies, although generative AI is spreading faster by several measures. Early adopters tend to have tasks where the immediate value is obvious and the necessary skills already exist. Later users arrive as interfaces become simpler, models support additional languages, prices fall, AI becomes embedded in existing products, and peers demonstrate practical uses.

Demographic averages can also conceal large within-group differences. Pew’s 2026 demographic analysis found that Asian American adults who responded in English reported higher chatbot usage than White, Black, or Hispanic adults in several measured categories. Sixty-six percent reported using chatbots for information search, and 60% of employed Asian adults reported workplace use.

Skill itself may compound adoption. Anthropic’s research suggests experienced Claude users achieve higher conversation success. OpenAI finds users attempt more distinct tasks as tenure increases. Employees who discover productive workflows can consequently gain more from subsequent model improvements because they already know where AI fits.

Governments have begun treating adoption skill as economic policy. New Space Economy’s discussion of Canada’s 2026 AI strategy illustrates this shift from research funding toward literacy, worker training, institutional adoption, compute access, and business deployment. Countries increasingly view the ability to use AI productively as separate from the ability to invent AI models.

The Central Pattern Is a Shift From Answers to Collaboration and Delegation

Across consumer, workplace, education, health, and technical research, one pattern connects much of the evidence. AI use tends to progress through stages.

A new user often begins by asking questions. The system substitutes for part of search, explanation, or reference work. Once trust and familiarity increase, users ask it to draft or transform material. They provide documents, data, images, or code and request outputs. More experienced users develop repeated workflows, provide richer context, connect outside information, and assign larger tasks. Agentic systems extend the process further by using tools and taking actions.

OpenAI’s consumer taxonomy captures the early division through Asking, Doing, and Expressing. Its 2026 enterprise data captures the next stage as companies move from assistance toward execution. Anthropic’s research records the same issue through augmentation and automation. The Anthropic Economic Index found that collaborative augmentation and more automated interactions continued to coexist, with the balance changing over time rather than moving uniformly toward complete delegation.

This changes the human contribution. When AI drafts an email, the employee supplies purpose, context, judgment, and authorization. When AI generates code, the developer decides what should be built and checks whether the result behaves safely. When a student uses AI as a tutor, the educational value depends partly on whether the student still performs the reasoning needed to learn. When a patient asks about symptoms, the system can supply information but cannot automatically substitute for examination, testing, diagnosis, or professional accountability.

AI use also expands the feasible task set. OpenAI’s July 2026 analysis of more than 800,000 work-related U.S. ChatGPT messages found that 16.8% concerned tasks associated with another occupation, and 43.5% of occupation-specific messages crossed conventional job boundaries. OpenAI describes this pattern as task crossover. A worker can use AI to perform limited programming, design, analysis, research, or communication tasks that previously might have required another specialist.

Anthropic’s April 2026 survey of 81,000 Claude users points in the same direction from self-reporting. Respondents in both the highest- and lowest-paid occupational groups reported substantial productivity gains, frequently because AI let them undertake work that had previously been outside their practical scope. People whose jobs were more exposed to AI also expressed greater concern about displacement, with concern higher among early-career workers.

That combination matters. AI can increase individual capability and increase concern about future labor demand at the same time. A worker may value the tool because it saves hours, then infer that an employer may eventually need fewer hours of the same labor.

New Space Economy’s analysis of the AI value chain helps explain why usage matters economically. Chips, data centers, cloud platforms, and models create capability, but applications capture value only when people place that capability into a task they want completed. Usage research shows where the infrastructure spending turns into actual behavior.

Summary

The strongest research available through August 26, 2026 shows that generative AI has become a general-purpose assistance layer spanning personal life, work, education, software development, medicine, search, creativity, and social interaction. No single use dominates every population. Information seeking, practical guidance, writing, learning, and content creation appear repeatedly across independent datasets.

Workplace AI is moving toward larger units of delegated work. Early adoption centered on drafts, summaries, research, and coding assistance. Enterprise systems increasingly connect models to files, tools, software environments, and organizational data so that agents can execute sequences of tasks. Software development is ahead of many occupations, but legal, sales, recruiting, marketing, finance, and administrative functions are moving in the same direction.

Education has reached mass adoption faster than most institutions have adapted their policies. Students use AI for explanations and brainstorming, but many also use it to obtain answers. Educators use the same systems for course preparation, administration, writing, and teaching tools. The educational issue has shifted from prohibition toward determining which forms of assistance support learning.

Health use now includes large volumes of personal information requests and widespread professional use by physicians. Emotional-support and companionship uses remain smaller among the adult population but are much more common among younger users, making them a distinct research and safety category.

Productivity evidence does not support one universal percentage gain. Controlled studies have measured substantial benefits in customer support, writing, and parts of knowledge work. Other experiments have found small gains, no gains, or temporary slowdowns. Self-reported productivity improvements tend to exceed experimentally measured effects. Task selection, user expertise, model capability, workflow design, verification burden, and organizational integration determine much of the difference.

The deeper change is behavioral. People are learning to divide work differently. Instead of asking whether AI can perform an entire occupation, actual usage shows tasks being split among humans and machines in changing proportions. Humans formulate goals, provide context, evaluate evidence, set constraints, review results, and retain accountability. AI handles growing amounts of searching, drafting, transforming, analyzing, programming, and execution.

The next measurement problem is no longer simple adoption. Counting users or chatbot visits will miss AI operating inside search, office suites, clinical software, coding environments, phones, educational services, and agents. Research will increasingly need to measure what portion of a task AI performs, how much human supervision remains, whether output quality changes, what new work becomes feasible, and whether the resulting time savings produce economic or social value.