The most common thing Americans do with an AI assistant is search for information. Not write code, not compose essays, not generate images. In Anthropic’s May 2026 Economic Index snapshot, the top work task matched to US Claude conversations is “search electronic sources, such as databases or repositories, or manual sources for information,” at 6.42 percent of task-matched usage, well above the 4.95 percent global figure. Pew’s February 2026 survey of 5,119 US adults lands in the same place from the opposite direction: among the 49 percent of Americans who use chatbots, searching for information is the single most cited use, at 42 percent.
Two more numbers reframe the standard story about American AI adoption. First, the United States ranks 12th out of 121 covered countries on per-capita Claude usage, behind eleven smaller economies, despite generating the largest absolute volume. Second, US usage skews personal: 49.7 percent of American conversations are personal rather than work or coursework, against a 40.2 percent global average. The country that builds most of the frontier models uses them, more than anywhere comparable, to plan trips, answer health questions, and manage hobbies.
This piece reads one primary dataset from top to bottom: what it measures, what it found nationally, how the 50 states and Washington, D.C. differ, and, most importantly, what the data cannot support. Every ranking page for “what do people use AI for” is a stat aggregator stitching together surveys with incompatible definitions. This is the opposite: one instrument, one snapshot, methodology first.
How the measurement works
The Anthropic Economic Index is built from anonymized, aggregated Claude conversations, not from asking people what they do. An automated classifier reads conversation content, with privacy aggregation preventing any tie back to individuals, and matches each conversation to the task it most resembles. The task catalog is O*NET, the US Department of Labor’s database that decomposes every occupation into its constituent tasks. A conversation about debugging a Python script maps to a software developer task. A conversation about drafting a patient education handout maps to a healthcare practitioner task. The method was introduced in the research paper Which Economic Tasks are Performed with AI? by Kunal Handa, Alex Tamkin, and colleagues at Anthropic, and the public dataset now extends it to 121 countries and every US state.
The geographic headline number is the Anthropic Usage Index: a region’s share of Claude usage divided by its share of working-age population. A value of 1.0 means usage exactly proportional to population. The US scores 3.87 against the global cohort, meaning Americans use Claude at nearly four times the rate their population share alone would predict, and still eleven countries score higher. Within the US, states are indexed against the national average instead, so a state’s 1.0 means typical for America, not typical for the world.
Two mechanical properties of this design matter for everything below. Task matching describes the task, not the person: a conversation matched to a nursing task tells you AI performed work of that shape, not that a nurse was typing. And a null in the dataset means not published, usually for privacy thresholds, never a measured zero. Survey data like Pew’s has the opposite character: it knows exactly who answered but relies on people accurately reporting their own behavior. The two instruments disagree in predictable ways, and where they converge, as they do on information search, the finding is unusually solid.
The national profile, including the surprises
The US topic mix breaks down as follows: content creation and copywriting leads at 20.0 percent, followed by hobbies and lifestyle at 13.6 percent, research and intelligence at 13.5 percent, education and learning at 9.3 percent, and software development at 8.1 percent.
The interesting information is in the deviations from the global baseline, and three of them cut against the prevailing narrative. Software development under-indexes in America: 8.1 percent of US usage against 11.5 percent globally. Content creation, the top category, also under-indexes, at 20.0 against 22.7. What over-indexes is the personal and advisory layer: hobbies and lifestyle runs 13.6 percent against 9.5 globally, and the artifact data shows the same tilt, with “advice or recommendation” as the output of 15.1 percent of US conversations against 10.7 percent worldwide. The stereotype of American AI usage as an army of developers shipping code describes the world’s usage better than America’s.
The occupational mapping sharpens this. Computer and mathematical tasks remain the largest single category at 21.1 percent of US usage, but that is below the 23.8 percent global share, an index of 0.89. The categories where America over-indexes hardest are sales and related tasks at 1.27 times the global rate, healthcare practitioner tasks at 1.31 times, and community and social service tasks at 1.25 times. The individual tasks behind those numbers are concrete: recommending products based on customer needs, advising clients on financial matters by phone or chat, and answering patients’ questions about their health each rank in the US top eight. American AI usage looks less like an engineering department and more like a service economy talking to itself.
The artifact data, which records what each conversation actually produces, completes the picture. The most common US output is an explanation or answer at 17.9 percent, followed by advice at 15.1 percent and a document or report at 13.9 percent. Emails and messages over-index against the world at 5.9 versus 4.6 percent, while apps and websites under-index at 3.3 versus 4.2. Line those outputs up and the modal American session ends with prose meant for a human, not code meant for a machine: an answer to read, a recommendation to act on, a document to send. Nearly 7 percent of conversations produce no artifact at all, which the taxonomy counts separately, and which likely covers the open-ended talk that surveys file under companionship.
One more national split deserves attention because it anchors most public argument about AI and work. The dataset classifies each conversation as augmentation, where the person iterates with the model, or automation, where the model completes the task with minimal back and forth. The US sits at 50.4 percent augmentation to 49.6 percent automation, statistically a coin flip, and close to the global 51.4 to 48.6. Whatever the future holds, the present, measured at the conversation level, is an even split between working with the machine and delegating to it.
Coursework is the sharpest US deviation of all: 8.9 percent of American usage against 16.5 percent globally. Barely half the world’s rate. Combined with the elevated personal share, the American profile is distinctive on both ends: less studying, more living.
Fifty-one profiles: the state data
The state-level table is where this dataset holds information available nowhere else, and the spread is larger than the national averages suggest. Washington, D.C. tops the index at 3.32 times the national per-capita rate, with the caveat that index rankings flatter small populations and D.C. is a city-sized unit packed with exactly the occupations that match well. Among actual states, California leads at 1.62, and California alone accounts for 19.25 percent of all US usage. New York follows at 1.55, Washington state at 1.51, Massachusetts at 1.42, and Colorado at 1.41. At the bottom sit Mississippi at 0.28 and West Virginia at 0.25. The distance from D.C. to West Virginia is a factor of 13. Even state to state, California’s per-capita rate is six and a half times West Virginia’s.
The profiles beneath the index diverge in ways the topline hides. Washington state, home to two of the largest software employers on Earth, posts the highest personal-use share among the top states at 54.3 percent, and only 36.2 percent work use. Florida is the mirror image: the highest work share among large states at 47.9 percent, the most automation-leaning large state at 52.7 percent, and a collapsed coursework share of 5.5 percent. Florida, on this data, is where Americans most often hand AI a work task and let it run.
Coursework is the most geographically variable split in the dataset. Iowa leads the country at 14.6 percent of usage, with education and learning as its number two topic statewide, a plausible college-town signature. Utah and Colorado sit near the bottom at 4.9 and 5.5 percent despite strong overall usage, pairing that with the most automation-leaning profiles among high-index states at 51.9 and 51.1 percent. High adoption and low student share travel together in the Mountain West, which reads as a workforce pattern rather than a campus one. Alaska bottoms out coursework entirely at 3.5 percent while posting one of the highest personal shares in the country at 54.9 percent.
The occupational rankings barely move across states, which is itself a finding. Computer and mathematical tasks rank first in all 51 profiles, from 18.5 percent in D.C. to 24.4 percent in South Dakota. What distinguishes states is the second tier: sales tasks climb to second place in Washington, Colorado, Utah, Minnesota, and Wisconsin, while D.C. and the Northeast tilt toward media, education, and administrative tasks. The American AI usage map is one national pattern with regional accents, not fifty different countries.
What this data cannot tell you
The limitations here are not fine print. They decide which conclusions survive, and the dataset’s own documentation is blunt about most of them.
First, this measures Claude, not AI. Pew’s survey puts ChatGPT at 44 percent of US adults and Google’s Gemini at 24 percent, so the dominant share of American AI activity flows through products this instrument cannot see. If Claude’s user base skews professional and technical relative to ChatGPT’s, and its enterprise positioning suggests it does, then the true American usage profile is likely even more personal and less code-heavy than these numbers show, because the largest consumer platforms are underrepresented, not the other way around. The direction of the headline findings survives the caveat. Their magnitudes should be held loosely.
Second, task matching is not occupation membership. When 4.3 percent of US usage matches healthcare practitioner tasks at 1.31 times the global rate, the defensible claim is that AI is being used for tasks commonly done by healthcare practitioners. Many of those conversations are patients, caregivers, and students, not clinicians. The dataset’s own methodology notes insist on this framing, and every secondary article that writes “nurses are adopting AI fastest” from this table has promoted the data past what it measures. Measurement categories quietly diverging from the phenomenon they get quoted as measuring is a pattern we keep meeting, most recently in the broken count of AI lawsuits, where a docket keyword and an actual AI dispute turned out to be very different populations.
Third, this is a snapshot with no trend series. The May 2026 period is one frame. Nothing here can show usage rising or falling, and nothing here can function as an early indicator of automation, job displacement, or labor demand. The dataset carries no income, employment, or productivity variables to correlate against, and the augmentation-automation split describes conversation style, not headcount. Anyone citing these numbers as evidence that a given occupation is being automated is making a claim the instrument cannot support, the same category of error as reading a benchmark score as generalization without checking what leaked into training, a failure mode we dissected in our taxonomy of data leakage in machine learning.
Fourth, the classifier itself is a model reading conversations, with the interpretive limits that implies. Task assignment at this scale cannot be hand-verified, and O*NET’s task catalog was written to describe human jobs, not chatbot sessions, so conversations that fit no occupational task get forced toward the nearest match. The published shares are best treated as a well-built estimate with unpublished error bars. How much to trust automated classifiers reading model behavior at scale is an open research question in its own right, one we covered this week from the mechanistic side in how researchers actually read an LLM’s mind.
Fifth, the per-capita index rewards small denominators. D.C.’s 3.32 is real arithmetic, but a compact, white-collar, government-and-media city will top any per-capita ranking of knowledge-work behavior. The more informative comparisons are between similar units: California against Texas at 1.62 versus 0.95 says something population size cannot explain.
What happens next
Three things are worth watching as this dataset matures. The first is longitudinality. The Index currently publishes snapshots, and the moment consecutive periods become comparable, the questions everyone actually cares about, whether coursework usage is growing, whether the automation share is creeping up, whether the state gap is closing, become answerable for the first time. A second frame turns a photograph into the beginning of a film.
The second is convergence testing against the survey record. Pew, fielded in February, and this telemetry, sampled in May, already agree that information search is the center of American AI usage and that work is not the majority use case. Where they diverge, the divergence is informative: self-report captures who, conversation data captures what, and the gap between them measures how badly people estimate their own behavior. Independent instruments agreeing from different directions is the strongest form of evidence this topic currently has. One axis this dataset cannot see, and the survey record can, is age. Pew’s age breakdown of the same survey shows about half of adults under 50 using chatbots to search for information, falling to 20 percent among those 65 and older, and one in five adults under 30 turning to chatbots for emotional support or advice, a share that halves for the 30 to 49 group. The anonymized telemetry aggregates all of that into single state and national figures. If the age gradient is as steep as Pew measures, then the personal-use tilt documented above is being pulled by a young cohort whose usage patterns will define the next decade of the profile, and the telemetry will register that shift only as a slow drift in topic shares with no explanation attached. The two instruments need each other.
The third is what the service-economy tilt does to product roadmaps. The measured American profile, heavy on sales tasks, financial advice, health questions, and recommendations, is a profile of advisory work, and advisory work is where errors carry consequences and regulators pay attention. A usage base that looks like a service economy pulls the industry toward reliability, sourcing, and accountability faster than a usage base of developers would, because developers can read a stack trace and a patient asking about symptoms cannot.
The honest summary of the data is narrower and stranger than the headlines built on it. America does not lead the world in AI adoption per person; it ranks 12th. Americans do not primarily use AI for work; the personal share is larger. The biggest usage category is not coding; coding under-indexes here. What the world’s largest AI market actually does with these systems, measured at the task level, is ask for information and advice about ordinary life, at four times the per-capita rate of the average covered country, split evenly between collaborating with the machine and delegating to it. That is the baseline any future claim about AI and American work should be tested against, and for one more period at least, it is a baseline with no trend line attached.