GenAI Reaches 80% of Occupations — but Most Workers Still Skip It

Three numbers describe AI at work in 2026: generative AI is now used in more than 80% of American occupations and in over 40% of detailed job tasks — and yet, within most of those tasks, fewer than half of the workers who perform them have adopted it. If you have been waiting for a definitive answer to "is everyone using AI except me?", this is it: almost everywhere, the answer is no. A new NBER working paper has, for the first time, measured genAI adoption not by occupation averages or platform anecdotes, but task by task, across the entire US economy.

[Fact] The paper — "What Work Does Generative AI Do?" by Alexander Bick (Federal Reserve Bank of St. Louis), Adam Blandin (Vanderbilt), David Deming (Harvard), and Tyler Schumacher (Vanderbilt), circulated as NBER Working Paper 35677 in August 2026 — draws on the Real-Time Population Survey (RPS), a nationally representative online survey of roughly 5,000 US adults per quarterly wave, with task-level genAI questions fielded across four waves from August 2025 to May 2026.

The first task-level census of what genAI actually does

The measurement trick is what makes this paper different. You cannot ask a survey respondent about thousands of possible work tasks. So the RPS first pins down each worker's detailed occupation under the 2018 SOC system, then shows them the ten most important O*NET work activities for that occupation, asks which they actually perform, and finally asks which of those they regularly use genAI to help with. That yields something no exposure index or chat-log study has produced: adoption rates for 2,040 detailed work activities, measured from workers themselves, with rankings that stay stable across all four survey waves.

[Fact] The headline levels, as of May 2026: 45% of US workers use genAI for work, 55% of adults aged 18-64 use it outside work, and 62% use it somewhere.

Widespread but shallow

The authors' own phrase for the pattern is "widespread but shallow," and the distribution backs it up.

[Fact] Over 80% of detailed occupations and over 40% of detailed tasks show adoption above 20% — but only about one in six occupations exceeds 70% adoption, just 2.8% of tasks clear 50%, and not a single task in the entire economy exceeds 70%.

[Fact] The top of the occupation table is a computing-and-analysis club: computer and information research scientists at 87.3%, information security analysts at 85.4%, computer programmers at 81.2%. The bottom is physical and interpersonal work: animal caretakers at 5.3%, receptionists at 7.6%, maids and housekeeping cleaners at 11.3%.

One quiet surprise sits in that top ten: chief executives adopt at 79.7% — a touch above software developers at 78.8%. The corner office is using genAI at the same rate as the people building it. If you work in one of the high-adoption fields, the detailed profiles for software developers, information security analysts, and computer programmers on this site cover how AI intersects with those roles.

Exposure scores predict less than half the story

Here is where the paper turns confrontational. Nearly everything written about AI and jobs over the past three years — including much of the analysis on this site — leans on "exposure scores": model-based predictions of which occupations AI could affect. This paper checks those predictions against measured adoption, and the scores come up short.

[Fact] Depending on the measure, exposure scores explain between 5% and 53% of adoption variation across occupations, and between 7% and 44% across tasks.

[Estimate] The authors' regression exercise puts it more bluntly: occupation exposure scores capture slightly less than half of the occupation-level variation in adoption that is available to be explained.

The misses are not random. [Fact] Receptionists are the single most over-predicted occupation: exposure-based models predict 54.0% adoption, while actual adoption is 7.6%. That prediction runs about seven times the observed reality — the model saw automatable information work; the workplace delivered compliance rules, thin technology training, and no incentive to experiment. Customer service representatives follow the same pattern, over-predicted by 31 percentage points.

The reverse misses are just as telling. [Fact] Construction equipment operators were predicted at 19.5% but actually adopt at 60.7%; special education teachers, chief executives, and computer repairers all beat their predictions by 25 points or more. What these under-predicted groups share, the authors note, is autonomy: nobody has to approve their prompts.

Your coworker matters more than your task

The paper's deepest finding is buried in a regression table. [Fact] Task fixed effects alone explain just 11.2% of the variation in whether a given worker uses genAI for a given task; adding individual fixed effects raises that to 44.0%. Do the arithmetic and the individual component explains nearly three times as much as the task itself. Demographics — age, sex, race, education — barely move the number.

In plain terms: knowing who you are predicts your AI use far better than knowing what you do all day. Two people doing the same job, side by side, routinely make opposite adoption choices.

[Fact] The authors trace part of this to learning that transfers across domains: workers with six-plus months of genAI experience use it for about 0.4 more tasks than newer users doing identical work, and instrumented estimates find that a 10-percentage-point rise in adoption across a worker's other tasks lifts adoption on any focal task by 5.4 to 7.0 points — with a similar spillover from work use into home use.

[Claim] The authors read this as a fixed-cost story: pay the learning cost once, in whatever domain offers a clear win, and applying genAI everywhere else gets cheaper — which implies adoption within occupations will keep deepening as experience compounds, even with no further model improvements.

A measurement warning aimed at chat-log studies

The paper's final section takes direct aim at the most-cited genAI datasets in circulation: task classifications built from platform chat logs — Anthropic's Claude data, OpenAI's ChatGPT data, Microsoft's Copilot data.

[Fact] The four measures barely agree. RPS task shares correlate with OpenAI's at 0.11, with Anthropic's at 0.34, and with Microsoft's at 0.10 — and the chat datasets do not even agree with each other, with every pairwise correlation below 0.4 and Anthropic-versus-Microsoft at 0.08.

The mechanism is over-classification into generic catch-all tasks. [Fact] More than 15% of OpenAI chats get classified as "Edit written materials or documents" — yet only 2.4% of US workers are in occupations that formally contain that task, and it accounts for just 0.1% of AI-assisted work in the survey. That is a 150-fold gap between what the chat classifier sees and what workers report. [Fact] Across the board, the top ten tasks absorb 46-61% of all use in chat data versus 22% in the survey.

Why it matters: chat-log task shares have been recycled into occupation-level "AI exposure" proxies across dozens of studies — and, as of last week, into official statistics. [Fact] The BLS's new observational AI exposure categories, covered in our analysis of the 2025-35 employment projections, are built partly on Claude and Copilot telemetry. This paper lands as a caution stamped on that entire measurement family: chat classifiers see the text of a conversation but not the occupation of the person typing, so they misattribute use across jobs.

Honesty requires two caveats pointed at ourselves. This site's occupation analyses draw on the same exposure-and-chat-log research family this paper critiques, so the correction applies here too. And the survey has its own blind spot: [Claim] the authors treat their numbers as a lower bound, since respondents cannot report embedded AI — autocomplete, meeting summaries, silent copilots — they never notice using, and an online panel may not perfectly represent offline workers even after weighting.

What this means for your career

Three practical readings. First, if you are a non-adopter in a high-adoption occupation, the risk is no longer hypothetical: the data show colleagues doing your exact tasks with genAI while you do not, and the learning mechanism means their lead compounds. Second, if you are in clerical work — the receptionists and claims processors the models over-predicted — low adoption around you is not proof of safety. It reflects barriers (compliance rules, data restrictions) that employers and regulators are actively working to remove; when they fall, adoption in your role may jump quickly, because the underlying work was always exposed. Third, the individual-variation finding is, oddly, the hopeful one: adoption is not fate assigned by job title. It is a learnable skill with demonstrated spillovers — starting anywhere, including outside work, measurably lowers the cost of starting everywhere else.

The deepest shift this paper makes is in the question itself. For three years the debate asked which jobs AI would change. The evidence now says most of the variation lives one level down: within the same job, between the person who adopted and the person who has not yet.

Sources

• Bick, A., Blandin, A., Deming, D.J., Schumacher, T. (2026). "What Work Does Generative AI Do?" NBER Working Paper No. 35677, August 2026. Also circulated as Federal Reserve Bank of St. Louis Working Paper 2026-017. nber.org/papers/w35677

This article was produced with AI-assisted analysis of the primary source. All figures were verified against the working paper's full text (62 pages, Federal Reserve Bank of St. Louis version dated August 25, 2026).