The file stores its own subtotals as ordinary rows. Average across them and you get a number that describes neither.
There is a row in Anthropic's Economic Index data that reads "15, Computer and Mathematical: 23.80." Elsewhere in the same file, in the same column, sit thirty-six rows for the individual computer and mathematical occupations: programmers, systems analysts, support specialists, web designers. Add those thirty-six numbers up and you get 23.79.
The file contains its own subtotals, stored as ordinary rows.
We know because we fell into the trap ourselves. Our first pass at this dataset computed a headline statistic across all 740 rows labeled soc_occupation: mean usage 0.2666, median 0.0200, a 13.3x gap between average and typical. Striking number. Also contaminated. The 740 rows are actually 718 detailed occupations plus 22 major-group aggregates, and the aggregates are sums of the details sitting beside them, so a naive average counts the same usage twice and inflates the mean by nearly double. Filter to the 718 real occupations and the mean drops to 0.137. The median does not move.
The corrected headline is a 6.9x gap between the average occupation's Claude usage and the median occupation's. That is still an enormous number, and everything in this piece is computed on the clean 718. But keep the trap in mind as you read, because it is a miniature of the essay's entire argument: this file physically stores the aggregate and the individual in one column, and averaging across them produces a number that describes neither. The mistake we are about to examine in AI-and-work commentary was waiting inside the data file itself, for whoever averaged first.
The Anthropic Economic Index is one of the only public, occupation-level measurements of real AI usage in existence. The release we analyzed (2026-06-26, on HuggingFace, window May 1 to June 1, 2026) maps Claude conversations to a taxonomy of roughly 17,000 ONET work tasks using Clio, Anthropic's privacy-preserving analysis system, then rolls task matches up to Standard Occupational Classification codes under ONET v30.2. The June release added hourly sampling, an artifact classifier, and a survey arm with about 9,700 respondents. It is serious measurement, seriously documented, and free.
Now the caveat that has to come before any conclusion, not after: this is Claude usage, not AI usage. The dataset cannot see ChatGPT, Copilot, Gemini, or anyone's in-house model. An occupation could be saturated with AI through another vendor and read zero here. There is academic work specifically on this platform-selection problem (Yin and Ogut, arXiv:2605.21743), and every sentence below should be read with "measured Claude usage" silently substituted wherever the prose gets loose. With that said: Claude is a large, general-purpose system, and there is no obvious reason its usage distribution would be wildly unrepresentative in shape even where it undercounts in level.
One more discipline item: units. The usage column is a share of observed task activity, expressed in percent. The median occupation's value of 0.02 means 0.02 percent of measured task-share, not two percent. Two orders of magnitude ride on reading that correctly.
Here is what the 718 detailed occupations look like, computed directly from the rows this week.
The mean usage share is 0.137. The median is 0.020. The ratio is 6.9.
But the mean-median gap undersells the shape. The distribution's real signature is at the bottom: 209 occupations, 29.1 percent of the entire detailed population, score exactly 0.00. Not low. Zero. Another tranche brings the count to 48.1 percent of occupations at or below 0.01, three-quarters (76.9 percent) at or below 0.10, and 96.8 percent at or below 1.00. The 10th percentile is zero. The 25th percentile is zero. The bottom quarter of the distribution is a flat line.
One honest asterisk on the zeros, and it is narrower than it was. Public releases of usage data sometimes suppress or floor very small cells for privacy, so the question is whether 0.00 here means measured-zero or below-a-disclosure-threshold. Two things bound it. Every one of the 740 rows carries a numeric value: there are no blanks, no nulls, and no suppression flags anywhere in the column, so a zero is something the release publishes rather than something it withholds. And a spot-check re-reading individual rows byte-for-byte against the full release file returns 0.0 as a stored value, not an absent cell.
That does not rule out flooring upstream, before publication, and we have not confirmed the point from Anthropic's methodology appendix. So the residual risk is real but specific: if a floor exists, "29.1 percent at zero" would partly measure a disclosure policy rather than the world. What survives either reading is everything the argument actually rests on. The median is 0.020, not zero. The 25th percentile is zero as published. And the typical American occupation, as measured by this release, has essentially no recorded Claude usage.
Put the whole distribution in one table and the shape stops needing description.
718 detailed occupations, Anthropic Economic Index release 2026-06-26. Bars are scaled to 100 percent.
share of occupations share of all measured usage
The top band is 3.2 percent of occupations and 44 percent of the usage. The bottom band is 29.1 percent of occupations and none of it.
| usage band | occupations | share of occupations | share of all usage |
|---|---|---|---|
| exactly 0.00 | 209 | 29.1% | 0.0% |
| 0.00 to 0.01 | 136 | 18.9% | 1.4% |
| 0.01 to 0.10 | 207 | 28.8% | 9.5% |
| 0.10 to 1.00 | 143 | 19.9% | 45.2% |
| above 1.00 | 23 | 3.2% | 44.0% |
Read the first and last rows together. The 209 occupations at the bottom hold none of the usage. The 23 at the top hold 44 percent of it. Those 23 are 3.2 percent of the occupations in the file.
Sit with that beside the discourse. Every week produces another essay about AI transforming work, and the essays are not wrong about the occupations they can see. But the modal occupation in Anthropic's own measurement is not lightly touched. It is untouched.
If the bottom of the distribution is a flat zero, the top is a spike. Compute the concentration and the numbers come out like this: the single highest occupation carries 5.2 percent of all measured usage. The top ten carry 26.4 percent. The top twenty-five, which is 3.5 percent of occupations, carry 45.9 percent. The top decile carries 71.4 percent.
The Gini coefficient of this distribution is 0.82. For calibration, no national income distribution on earth is that unequal; 0.82 is the concentration regime of wealth, not of earnings. Claude usage across occupations is distributed like assets, not like salaries: a small class holds most of it, and the median holder has approximately none.
Who is the top? Here are the ten highest-usage detailed occupations in the release, verbatim from the rows:
| usage | occupation |
|---|---|
| 5.10 | Document Management Specialists |
| 4.22 | Librarians and Media Collections Specialists |
| 2.72 | Counter and Rental Clerks |
| 2.58 | Editors |
| 2.22 | Writers and Authors |
| 2.17 | Computer User Support Specialists |
| 1.91 | Computer Systems Analysts |
| 1.70 | Web and Digital Interface Designers |
| 1.67 | Computer Programmers |
| 1.67 | Technical Writers |
Mostly this list confirms priors: text-heavy and code-heavy work dominates. Two entries deserve more attention. The leader is not a software job; Document Management Specialists at 5.10 is the single largest cell in the file, which makes a certain sense (the job is nearly all text handling) and still surprises. And third place is genuinely strange: Counter and Rental Clerks, a retail occupation, above Computer Programmers. We flag it rather than explain it. It may be real (correspondence, scheduling, inventory queries), or it may be a task-mapping artifact, since Clio assigns conversations to O*NET tasks and generic clerical tasks are shared across many occupations. An anomaly honestly labeled is worth more than a story invented to cover it.
The family-level comparison makes the concentration vivid. The Computer and Mathematical family is 36 occupations, five percent of the population; it carries 24.1 percent of all measured usage, and its median occupation sits at 0.255, about thirteen times the overall median. Healthcare Practitioners and Technical is the largest family in the dataset, 77 occupations; it carries 3.3 percent of usage, with a median of 0.010, half the overall median. The largest occupational family in America's labor taxonomy is in the untouched majority. That is the thesis of this piece in one comparison, and it is computed, not asserted.
The release includes, for 716 of the 718 occupations, a split of usage into automation-flavored interactions (the model does the task) and augmentation-flavored ones (the model helps a human do the task). The splits sum to 100 per row. Averaged across all detailed occupations, automation edges augmentation, 51.7 to 48.3.
Now split the population by usage level. The twenty heaviest-using occupations average 49.0 automation and 51.0 augmentation: the occupations that use Claude most lean toward augmentation. The 399 occupations at or below the median usage average 55.2 automation: the occupations that barely use it at all lean toward automation.
The popular narrative runs exactly backwards from this. Heavy AI adoption is assumed to mean replacement, light adoption safety. In this release, the heavy users skew collaborative and the near-nonusers skew substitutive. A caveat must travel with the finding: the bottom group's percentages are computed on tiny usage denominators and are noisy by construction, so the claim is directional, not precise. And the direction moves across releases; Anthropic's own reporting found 57 percent of usage augmentative in its first report, with automation increasingly prevalent by mid-2026. But the cross-sectional shape in this release is the opposite of the folk model, and it deserves to be on the record.
A finding this stark should make you suspicious of the instrument, so it matters that Anthropic's own task-level reporting says the same thing through a different denominator. In the June 2026 report, the top ten O*NET tasks account for 24 percent of all measured usage, up from 21 percent in January 2025. Tasks and occupations are different units, computed by different people, and both come out heavily concentrated, with concentration rising. Their reporting also notes that the tasks AI covers require 14.4 years of education on average against an economy-wide task average of 13.2, meaning the usage is concentrated in the more schooled portions of work.
Concentration is not a quirk of our arithmetic. It is the structure of the phenomenon.
The physicist Ole Peters built a research program (the canonical paper is in Nature Physics, 2019) around a simple observation with deep consequences: the average across an ensemble and the experience along a single trajectory can diverge, and economics habitually reports the first while people live the second. His policy recommendation is disarmingly modest: report both.
Strictly speaking, our dataset supports the older, simpler cousin of that argument. We have a cross-section at one moment, not trajectories through time, so what the numbers demonstrate is that a heavily skewed distribution makes the mean unrepresentative of the typical case, which statisticians have preached forever. We use the ergodicity frame as a lens, not a proof. But the lens focuses the right thing: "average AI exposure" statistics describe an ensemble in which Document Management Specialists and dental hygienists are melted together, and no actual occupation lives at the average. The mean occupation experiences 0.137. The median occupation experiences 0.020. The modal occupation experiences zero. When someone says AI is transforming work, the honest follow-up question is: at which percentile?
Three practical takeaways, one per audience.
If you analyze data for a living: when a file stores multiple aggregation levels under one category label, the mean of the column is a landmine, and the file will not warn you. Check whether subsets sum to other rows before averaging anything. Our 13x-versus-6.9x embarrassment is reproducible in twenty lines of Python, and so is the check that catches it.
If you build or sell AI products: the top decile, seventy percent of current usage, is where competition already lives. The interesting territory is the middle half, at 0.01 to 0.10, where usage exists but is thin. Moving an occupation from 0.02 to 0.2 is worth ten times more, in distribution terms, than another point of share among programmers.
And if you make or consume claims about AI and the economy: insist on both numbers. The mean says transformation. The median says 0.02 percent. Both are true, they are seven-fold apart, and the gap between them is the actual state of AI and work in mid-2026.
release_2026_06_26/data/aei_claude_ai_2026-06-26.csv), category_name=soc_occupation, geo_level=global, window 2026-05-01 to 2026-06-01. All distribution statistics in this piece (718/22 row split, mean 0.137, median 0.020, 209 exact zeros, Gini 0.82, concentration shares, family aggregates, automation/augmentation splits) computed directly from the rows on 2026-08-04; the aggregation check (detailed family sums vs major-group rows, e.g. 23.79 vs 23.80 for SOC 15) computed the same day.Which percentile is your number from?
The mean and the median of this dataset are seven-fold apart, and a claim that reports only one of them describes a distribution nobody inhabits. The same problem appears wherever an aggregate stands in for a trajectory, including in how AI agents report their own work: a summary averages over everything the agent did, when the thing you need is the specific path it took. Chain of Consciousness records that path, a tamper-evident log of what an agent did, on what inputs, in what order, written as the work happens rather than summarised afterwards. The trajectory instead of the ensemble.
Hosted Chain of Consciousness · Verify a record
pip install chain-of-consciousness · npm install chain-of-consciousness