Methodology

This page documents every method, source, and known weakness in the atlas. It is written to be checked: each statistic on this page traces to a file in the public data bundle or the open pipeline repository, and the limitations section is complete to the best of our knowledge. If you find an error, the contact address is on the About page.

Summary

The Colorado AI Exposure Atlas estimates, for each of Colorado’s 64 counties, how much of the local workforce holds jobs whose tasks overlap substantially with AI capabilities assessed in 2023. It combines county employment records (BLS QCEW, 2025 annual averages) with national occupation staffing patterns (BLS OEWS, May 2025) and published occupation-level AI exposure scores (Eloundou et al., Science, 2024). Statewide, the employment-weighted exposure score is 31.4 on a 0–100 scale, and 27.4 percent of UI-covered jobs are in occupations where at least a large minority of tasks overlap with AI capabilities assessed in 2023 — the top two of the four exposure tiers this atlas reports, occupations scoring 45 or above. The tier cuts are a choice on a continuous score; the exposure-measure section defends the choice and shows what the alternative cut would report. Exposure is not a job-loss forecast: it measures where AI-driven change — augmentation and automation both — is likely to arrive first. County occupation mixes are estimated from industry mix, not observed, and every county page reports how much of its underlying data was imputed.

What exposure means (and doesn’t)

An occupation’s exposure score measures one thing: the overlap between the tasks that make up the job and AI capabilities as assessed by human annotators in 2023.

It does not distinguish augmentation from automation. A task can overlap with AI capability because AI could do it instead of the worker, or because AI could help the worker do it faster. The score counts both. In one week of consumer conversations, model-classified usage (see the observed-usage section below) had augmentation at least as common as automation — a thin base for a strong claim in either direction.

It does not predict job loss. Whether exposure becomes displacement depends on adoption costs, regulation, task reorganization, labor markets, and choices firms haven’t made yet. None of that is in this data. What the atlas can defensibly claim is narrower: it maps where in Colorado the change — whatever form it takes — is likely to arrive first, because that is where the task overlap is concentrated.

It also does not measure current AI adoption. An accounting-heavy county scores high whether or not any accountant there has touched an AI tool. Exposure describes what could change; whether it is changing is an adoption question, which enters this atlas only through the national usage data described below.

What this can and can’t tell you

What the atlas measures is task overlap: how much of each occupation’s work, as catalogued in 2023, overlaps with AI capabilities assessed in 2023, aggregated over each county’s estimated occupation mix. It is not a forecast: it does not estimate displacement or predict net employment.

The strongest criticism of this genre belongs on this page. Benedict Evans made the case at length in May 2026 (Predicting AI job exposure): you do not know how jobs will change, you do not know what will change around them, and work is not reducible to a task list in a database. The track record supports the caution. The best-known earlier index — Frey and Osborne’s 2013 automation-risk estimates — classified nearly half of U.S. employment as at high risk of computerization; a decade on, no displacement on that scale had appeared. Exposure indices have been more reliable as maps of task overlap than as predictions of outcomes, and this atlas should be read as the former.

Method: how county scores are built

Establishment-based county occupation data does not exist. The Bureau of Labor Statistics publishes county employment by industry (QCEW) and occupation employment only down to the state and metro level (OEWS). The Census Bureau’s ACS does publish county occupation-group estimates, but those are household-reported and residence-based — they count people rather than jobs, stop at broad occupation groups, and carry large margins of error in small counties. Because they measure where workers live rather than what county workplaces employ, they are not the input here (though they are a candidate cross-check for a future edition). So the atlas imputes each county’s occupation mix from its industry mix, using national staffing patterns. For each county c:

emp(c, o)    = Σ_i  emp(c, i) × share(o | i)

exposure(c)  = Σ_o  emp(c, o) × score(o)  /  Σ_o emp(c, o)

where:

• emp(c, i) is the county’s 2025 annual-average QCEW employment in industry i, taken at the finest disclosed NAICS level (4-digit where published, falling back to 3-digit or sector), with suppressed cells imputed as described below;
• share(o | i) is the national share of industry i’s employment in occupation o, from the OEWS May 2025 national industry staffing matrix (detailed occupations only, renormalized to sum to 1 within each industry);
• score(o) is the occupation’s Eloundou et al. human-β exposure rating, scaled 0–100.

The county’s exposure score is the employment-weighted mean over its implied occupation employment. The headline high+substantial share is the fraction of that implied employment in occupations scoring 45.0 or above — the top two of the four tiers defined in the exposure measure — and allocating the same implied employment across all four tiers gives the county’s tier distribution. The secondary share published alongside it, retained from earlier editions, is the fraction scoring 50.0 or above. Both QCEW 2025 and OEWS May 2025 are coded to NAICS 2022, so the industry join is direct, with no vintage crosswalk.

This industry-mix imputation is the established approach for county occupation estimates, not an invention of this project. A widely cited county-level exposure study — Brookings’ Automation and Artificial Intelligence: How machines are affecting people and places (Muro, Maxim & Whiton, 2019) — used county occupation estimates purchased from EMSI, whose county estimates are built primarily on OEWS staffing patterns applied to county industry employment. This pipeline is an in-house, open-source equivalent of that vendor step. Brookings’ October 2024 report Generative AI, the American worker, and the future of work (Kinder et al.) likewise uses the Eloundou β measure aggregated with OEWS employment — the same primary measure used here. We cite the lineage for a narrow reason: the method’s known weaknesses (see Limitations) are the field’s weaknesses, not shortcuts specific to this atlas.

What we do not publish: detailed-occupation estimates for individual counties. The imputation is defensible for a county’s aggregate score and its occupation-group profile; it is not strong enough to defend a claim like “Pueblo County has 340 paralegals.” County occupation detail stops at the SOC major group (22 groups), and we would treat any finer figure derived from our files with suspicion.

Sources

Every input dataset, with vintage and license. Raw files are fetched by the pipeline’s fetch.py and verified against committed checksums.

A note on the Felten AIOE data. The AIOE repository publishes no license, which under copyright default means all rights reserved. We therefore never republish the raw AIOE values. The pipeline fetches them at build time, and the atlas reports only derived statistics — the correlations in the robustness section. Anyone reproducing our build fetches the AIOE file from its authors’ repository directly.

The exposure measure

The primary score is the human-annotated β measure from Eloundou et al. Annotators classified each O*NET task as E1 (an LLM alone could cut the time to do it well by at least half), E2 (the same, but requiring additional software built on the LLM), or E3/E0 (neither). An occupation’s β is the share of its tasks rated E1 plus half the share rated E2. We scale it by 100, giving a 0–100 score where 100 would mean every task is directly LLM-exposed. The observed 2025 range across occupations is 0 to 84.4.

Both ends of that range deserve a note. At the bottom, 55 of the 830 published occupations score exactly zero. We checked each against the raw Eloundou file: every contributing ONET variant carries a literal 0.0 rating — genuine annotations, not missing values coerced to zero — and the roster is overwhelmingly physical-task work (26 of the 55 are construction and extraction occupations). The site renders these with the tier language, “few tasks overlap,” rather than the numeral, because “0.0” states a 2023 annotation with more certainty than it carries; the exact values are in the data downloads. At the top, two occupations — Animal Scientists and Survey Researchers — share the maximum score of 84.4 because their raw ratings are identical, exactly 27/32 in each case, each from a single distinct ONET variant: a verified coincidence of the source ratings, not a join artifact.

Why β: it is the paper’s middle measure — broader than α (E1 only, direct exposure), narrower than ζ (full credit for E2). It is the measure Brookings adopted for its 2024 generative-AI work, which keeps our figures comparable to the national literature. Why the human columns rather than the paper’s GPT-4-generated ratings: using a model to grade its own exposure invites an obvious circularity objection, and the human ratings avoid it. (The two agree closely in the paper; we simply don’t need to lean on that.)

The scores cover 798 of 867 SOC 2018 detailed occupations. The 69 without scores are 19 military occupations (excluded — military employment is not UI-covered and is negligible in QCEW) and about 50 residual “all other” catch-all codes, which are assigned their SOC major group’s national-employment-weighted mean score. Where multiple O*NET variants map to one SOC code, we take the unweighted mean, matching standard practice.

The exposure tiers. Since v1.2, each occupation sits in one of four tiers by its score, with inclusive lower bounds at 60, 45, and 30: high — “most tasks overlap with AI capabilities assessed in 2023” — at 60 or above (52 of the 830 published occupations); substantial — “a large minority of tasks overlap” — at 45 to 59.9 (151 occupations); some — “some tasks overlap” — at 30 to 44.9 (195); and little — “few tasks overlap” — below 30 (432). The cuts are round numbers on a scale that already has a reading — the score is a weighted share of an occupation’s assessed tasks, so a tier boundary at 45 is a statement about tasks, not a quantile of an employment distribution that shifts with every edition. The half weight on E2 tasks carries the same caveat as before: E2 exposure is conditional on tooling actually being built on top of the model, so every tier boundary is generous in that respect. Statewide, the four tiers hold 6.3, 21.1, 27.9, and 44.7 percent of UI-covered employment respectively.

The headline is the top two tiers. Where the atlas quotes one number, it is the high+substantial share: the fraction of employment in occupations scoring 45 or above — 27.4 percent statewide, 25.9 percent nationally. Any cut on a continuous score is a choice, this one included. The tiers do not remove the arbitrariness of a threshold; they publish it at three boundaries a reader can inspect and move.

Why the binary was replaced. Through v1.1 the atlas reported a single cut: an occupation was “high-exposure” at a score of 50.0 or above, the top quintile of the national employment-weighted score distribution. A hard cut at the midpoint of a continuous measure manufactures distinctions the source data does not support. The clearest case in the Colorado data: Software Developers score 44.7 and Computer Programmers 68.3, so the binary counted all of Colorado’s roughly 43,000 software developers as not exposed and all of its roughly 900 computer programmers as exposed. The new cut does not dissolve the problem — 44.7 sits 0.3 below the 45 boundary too, and a 44-versus-45 choice of boundary would move those 43,000 jobs across the headline. That is why the site now reports a distribution over four tiers rather than a single share, and why the boundaries are published: a reader who would draw the lines elsewhere can recompute every figure from the data files.

A footnote on that pair, since it invites the question of whether the 23.6-point gap is real. Both figures come directly from the Eloundou et al. human-β ratings for the single ONET occupation feeding each SOC code — exactly 17/38 for developers (0.447), 41/60 for programmers (0.683) — with no crosswalk aggregation or imputation involved. The two SOC codes are distinct occupations with different ONET task lists, and the ratings are ratings of those task lists, not of job titles. Our inputs carry only the occupation-level β, so the atlas reports the difference without decomposing it.

What the old cut would show. Both cuts are published, so the choice can be judged rather than taken on faith. Under the previous cut — 50.0, the top quintile of the national employment-weighted score distribution — 23.1 percent of Colorado’s UI-covered jobs are high-exposure, against 21.7 percent nationally. That national figure is close to definitional: the threshold was defined as the national top quintile, so the national share lands near a fifth by construction, and only the state-minus-national gap carried information. (It also carries a tie: 16 occupations sit at exactly 50.0, straddling the quintile boundary; counting them in, as we always have, puts 22.0 percent of national OEWS employment above the cut rather than 19.1.) Under the v1.2 cut — 45, a fixed point on the score scale rather than an employment quantile — the shares are 27.4 percent statewide and 25.9 percent nationally, and the national figure is a measurement of the national employment mix, not an echo of the threshold’s construction. The ≥50 shares remain published for every county under their original column names.

How counties are presented: shares, bands, and anchors

Added in the v1.1 revision (August 2026); recut in v1.2 on the tier headline. Everything in this section is presentation; no underlying value changed.

The headline statistic is a share of jobs. Through v1.0 the county pages led with the 0–100 mean score. Since v1.1 they lead with a share of UI-covered jobs — stated as a fraction and a count of jobs — with the mean score demoted to the tables, the data files, and this page; since v1.2 that share is the high+substantial share defined in the exposure-measure section. The reason is resolution. County mean scores span 24.7 to 35.2 — 10.5 points of a 100-point scale, median 28.9 — and ranking 64 counties on so narrow a spread manufactures distinctions the method cannot support. Kiowa and El Paso counties both publish a mean score of 31.4 and hold ranks 7 and 8, separated at a decimal we do not print. A share of a workforce needs no scale to interpret. In page prose, job counts are rounded to the nearest hundred; exact counts stay in the tables and data files.

Map colors. The job-sized view and geographic map share one color scale and legend, referenced to Colorado’s employment-weighted statewide share. Switching the view changes the size and shape of counties, not their colors. The typical-county bands described below remain available as text-only labels in the detailed table; their colors no longer compete with the map scale.

Bands. Each county sits in one of five bands by its high+substantial share: even quintiles by county count — 13 counties in each of the top four bands, 12 in the bottom — cut on the full-precision share, remaining ties broken by FIPS code. The labels: well above the typical Colorado county, above the typical county, about typical, below the typical county, well below the typical county. Because the cuts use full precision, two counties can display the same rounded share and sit in different bands; the v1.1 banding had such a pair (Phillips and Prowers, both displaying 16.5 percent), and the v1.2 recut on the new headline happens to have none — a property of this edition’s data, not of the method. This edition’s band edges are in the data files.

A footnote on the cut method. The bands could instead have been cut so each held a fifth of state employment. The edges would land elsewhere: employment is concentrated in high-share metro counties, so employment-weighted cuts would isolate a few large counties in the upper bands and sweep most of the state’s counties into the lower ones. We chose count-based cuts because they answer the question a county page asks — where does this county stand among counties — and because “13 counties per band” can be checked by counting.

Why the labels anchor to the typical county, not the state average. The statewide share, 27.4 percent, is employment-weighted, and it sits above the individual shares of 58 of the 64 counties, because most Colorado jobs are in a handful of metro counties whose shares run high. Labels anchored to the state figure would mark nearly every county “below average” and tell a reader nothing. So the labels anchor to the typical (median) county — a share of 21.2 percent — and comparison graphics draw the statewide figure as its own marked line. The gap between the two numbers is a finding in its own right: most Colorado jobs, and most exposed jobs, are in the same handful of metro counties.

Ranks, and when we withhold them. The 1-to-64 rank remains published for every county in the table and the data files. It no longer appears in county-page prose, meta descriptions, or link-preview text for any county with more than 30 percent of its employment imputed — 38 of the 64. For those counties the ordinal is the most quotable and least defensible number the method produces (see the Crowley question in the FAQ); the band and the data-quality badge carry the comparison instead.

Peer counties. Each county page names the three other counties nearest to it in total employment (nearest by log-employment distance), with their shares, so a reader can compare like-sized economies rather than every county against Denver.

The national anchor. The atlas publishes a national share for each cut, computed by the pipeline that produces every county figure — national QCEW industry employment routed through the same OEWS staffing patterns, the same exposure scores, the same cuts — so county, state, and national shares are one unit measured one way. For the v1.2 headline the anchor is 25.9 percent: the share of national UI-covered employment in high+substantial occupations. Because the 45 cut is a fixed point on the score scale, this one is a measurement of the national employment mix — the number a county’s share can fairly be compared against. The v1.1 anchor of 21.7 percent, under the ≥50 quintile cut, remains published with the caveat it always carried: that threshold was defined as the top quintile of national employment ranked by exposure score, so its national share lands near a fifth by construction. (It reads 21.7 rather than the 22.0 percent in the exposure-measure section because the employment base differs: the anchor uses QCEW UI-covered employment routed through staffing patterns, where the threshold is defined over OEWS occupation employment directly.) The anchors are the denominators that let a county page say “20.6 percent here, 27.4 statewide, 25.9 nationally” in one like-for-like unit.

For an outside reference point: Brookings’ October 2024 generative-AI report (Kinder et al., cited in the method section above) also reports exposure as shares of jobs, and finds metro-area shares above nonmetro shares — roughly 35 versus 30 percent on its threshold. Its threshold and score treatment differ from ours, so the levels are not comparable and we do not use them as a benchmark. What the comparison supports is direction only: an independently built measure also finds exposure concentrated in metropolitan employment, which is what this atlas finds within Colorado.

The occupation pages

Also new in v1.1: a page for each of 830 detailed occupations, under /occupation/. The county estimates in this atlas rest on imputation; the occupation pages do not use county-level imputation. Each page joins two published datasets — the occupation’s Eloundou human-β exposure score, the same measure used everywhere on this site, and the BLS OEWS May 2025 Colorado state file’s published employment and median wage for that occupation. Employment and wages are published BLS estimates; any occupation score filled from its major group is explicitly labeled. Where BLS publishes no Colorado value, the page says “not published” rather than filling the gap: 718 of the 830 occupations carry a published Colorado employment count, and 741 a Colorado median wage. Three published-data conventions: 59 of the 830 scores are the major-group fill described in the exposure-measure section (residual “all other” codes), and each such page says so; the four Colorado median wages BLS top-codes are shown at the published threshold of $239,200; occupations BLS reports only hourly are annualized at 2,080 hours. Read that $239,200 as a floor rather than a median. Those four pages mark the threshold with ≥ and explain that it is a reporting floor. It is not a ceiling: BLS publishes Colorado medians above $239,200 for 13 occupations — twelve medical or dental specialties and airline pilots — and those pages carry the published figure. The percentile strip in each page’s expandable technical details is the share of the 830 scored occupations with a strictly lower score, unweighted by employment. Full conventions are in the data bundle’s README.

The illustrative occupation lists on county pages — the named occupations under each occupation-group bar — come from the same published national data: the six largest detailed occupations in each group by national employment, excluding residual “all other” codes. They illustrate what a group contains nationally. They are not county estimates, and publishing them does not breach the rule against detailed-occupation county figures (see “What we do not publish” above).

Robustness: does the choice of measure matter?

A fair objection to any exposure atlas is that the author picked the measure that told the best story. We checked the obvious alternative: the Felten, Raj & Seamans AI Occupational Exposure (AIOE) index, built from a different method (linking AI capability benchmarks to occupational abilities rather than annotating tasks) in a different year (2021).

• At the occupation level, Eloundou β and AIOE rank occupations with a Spearman correlation of 0.855 (across 772 crosswalked occupations).
• At the county level — running the full county pipeline with AIOE substituted as the score — the Spearman correlation between the two county rankings is 0.934.

The county ranking barely depends on which measure is chosen. The pipeline’s validation gate fails the build if this correlation drops below 0.7. Because AIOE values are unlicensed, they appear in the atlas only as these derived statistics; the crosswalk from AIOE’s SOC 2010 coding to SOC 2018 is many-to-many, and for the 22 of 772 mapped codes where several 2010 values merge, we take their unweighted mean.

Observed usage: the Anthropic Economic Index

Exposure scores are predictions about task overlap. The Anthropic Economic Index (AEI) offers something different: observed usage of one AI system (Claude), classified by task and occupation. We use it as a third lens on the methodology page only — AEI data never enters the county scores — to check whether predicted exposure and observed use point the same way.

• Predicted vs observed, nationally: ranking the 22 SOC major occupation groups by predicted exposure (employment-weighted Eloundou β) and by observed share of U.S. Claude.ai usage gives a Spearman correlation of 0.74. The agreement is uneven. Computer and mathematical occupations agree strongly: predicted most exposed of the 22 groups, and first in observed usage at roughly 26 percent of classified conversations. The divergences are just as visible: Business and Financial Operations is predicted second-most exposed but accounts for only about 3.4 percent of observed usage — sixth among the groups, below Education and Arts — and Legal, predicted fifth-highest, shows about 0.5 percent of observed usage.

  Prediction and observation can diverge this way without either being wrong, because they measure different things. Exposure scores task overlap; observed usage adds adoption friction (professions differ in how quickly they pick up new tools and in what their rules, clients, and confidentiality obligations permit), a consumer-vs-enterprise split (the data covers consumer Claude.ai conversations, so a student’s essay help counts and a law firm’s contract-review platform does not), and the gap between how many of an occupation’s tasks could involve AI and how many conversations those tasks actually generate.

• The feasibility check: 97.17 percent of observed usage (occupations weighted by their observed-usage measure in the AEI labor-market files) falls in occupations whose Eloundou human α rating — the share of tasks directly LLM-exposed — is greater than zero. Under the broader β criterion the share is 99.82 percent. This is an occupation-level statement: observed usage overwhelmingly lands in occupations the exposure framework already identified as having some directly exposed tasks. The AEI companion files do not carry task-level Eloundou ratings, so we do not restate this at the task level.

• Colorado specifically: Colorado’s per-capita Claude usage index is 1.30 — eighth of the 51 states-plus-DC — and in one week of consumer conversations, model-classified usage tilted slightly more toward augmentation (54 percent augmentation / 46 percent automation) than the U.S. baseline (51/49).

Caveats on the AEI data, all of which are reasons we keep it out of the county scores: it observes Claude conversations only, not other AI systems; the state-level data covers consumer Claude.ai (Free and Pro) only, excluding enterprise API traffic; geography is IP-geolocated; the state file is a one-week snapshot (August 4–11, 2025); privacy thresholds suppress small cells, so a missing value means unreported, not zero; and the task/occupation classification is itself model-judged. We use the September 2025 geographic release because it is the most fully documented state-level snapshot available at build time — later releases restructure the reported facets — and will evaluate the 2026 monthly releases for the 2027 edition.

Suppression and imputation

QCEW suppresses county × industry cells small enough to risk identifying individual employers. Suppression is unavoidable in county-level work; the rules below state exactly how we handle it.

The imputation rule. Within each county, the employment in suppressed cells is recoverable in total (disclosed parent totals minus disclosed children) but not in distribution. We distribute each residual across its suppressed sibling industries in proportion to the state-level employment mix of those industries, honoring floors implied by disclosed children. Where the state-level cell is itself suppressed (6 cells at the 3-digit level, 29 at 4-digit in 2025), the state pass uses an equal split among suppressed siblings. Employment is never dropped: each county’s occupation employment is asserted to sum to its QCEW total within 0.01 percent. The direction of the resulting error is predictable: distributing by state-level mix pulls a heavily-suppressed county toward the state’s industry composition, so a small county whose true mix is genuinely unusual — in either direction — is likely reported as less unusual than it is. Extreme small-county scores are best read as understated at both ends.

How much is imputed, county by county. Statewide, 8.6 percent of employment sits in imputed cells, but the range across counties is wide: from 1.8 percent in Denver County to 86.4 percent in Gilpin County (San Juan County, with 448 covered jobs, is at 75.4 percent). Every county page prints its exact pct_imputed figure, and the full table is in DATA_NOTES. Alongside the number, each county carries a data-quality badge: full (under 10 percent imputed — 9 counties), partial_imputed (10–50 percent — 38 counties), or heavily_imputed (over 50 percent — 17 counties), with heavily-imputed counties hatched on the map. The bands (decided August 6, 2026, after reviewing the actual distribution) drive the badge and the map hatching — moving a band edge would change which counties are hatched. The published pct_imputed figures do not move with the edges, and are always shown alongside the badge.

Fallback routing. A separate, smaller issue: some QCEW industries have no published OEWS staffing pattern, so their employment must borrow one. The rules, in full: public administration (NAICS 92xx) is routed to a blend of the OEWS state-government and local-government staffing patterns (weighted by national employment; the federal pattern is deliberately excluded, because with QCEW ownership detail aggregated away there is no defensible county-level federal weight, and county 92xx employment is dominated by state and local government). Agricultural industries OEWS doesn’t cover are routed to the crop- and animal-production support-activities patterns — the nearest covered farm work, not the all-industry average. Private households and unclassified rows take the all-industry national pattern. Fallback-routed employment totals 6.5 percent statewide and is reported per county in DATA_NOTES.

Validation

The pipeline enforces seven gates on every build; if any fails, no data is emitted. The observed 2025 results:

1. County-to-state QCEW total. The 64 county totals sum to 2,860,276 jobs against a published state total of 2,891,093 — a gap of 1.07 percent. The gap has an identified source: QCEW books 30,991 Colorado jobs to an “Unknown or Undefined” pseudo-county (area 08999) that no county file can contain, which accounts for all but 174 jobs of the observed 30,817-job difference — a residual of 0.006 percent of the state total. The atlas reports statewide figures as the sum of counties, so that county pages always add up to the state figure we cite, with this note attached.
2. Completeness and ranges. All 64 counties present in every emitted file; every published number non-null; all scores within [0, 100]; county ranks a permutation of 1–64; slugs unique.
3. OEWS reconciliation. The strongest external check available: our imputed statewide occupation distribution, compared against the occupation distribution BLS actually publishes for Colorado (OEWS state file), as shares by SOC major group. Across the 22 groups, the mean absolute difference is 0.57 percentage points and the maximum is 2.35 points. The maximum is an offsetting pair — Management overcounted by 2.06 points, Business and Financial undercounted by 2.35 — consistent with a classification difference between survey instruments (where the line between “manager” and “business specialist” is drawn); restricting the comparison to occupations published on both sides moves the maximum only to 2.28. We cannot rule out a genuine difference between Colorado’s occupation mix and national staffing patterns — this is precisely the kind of deviation the method cannot see. The next-largest gaps are Sales (−1.62) and Office and Administrative Support (+1.08). The full 22-row table is published in DATA_NOTES. The gate’s tolerances — 2.5 points maximum, 0.75 mean — were set after this baseline was observed, chosen with the 2.35/0.57 results in hand, so the 2026 build could not have failed them. They are frozen for future editions, where they become binding.
4. Eloundou-vs-AIOE Spearman. The county rank correlation between the primary (Eloundou) and robustness-check (AIOE) scores — 0.934 — must exceed 0.7.
5. Determinism. The full pipeline, run twice, must produce byte-identical output files. It does, on every build.
6. Committed outputs current. The data files committed to the repository must match a fresh build byte-for-byte; a stale committed dataset fails the build.
7. Derived presentation fields. Added with v1.1 and extended for the v1.2 tiers: band sizes and ordering against the headline share (13/13/13/13/12), peer-record integrity, every published job count consistent with its share times employment within rounding, the national anchors inside sanity bounds, occupation-file coverage of the scores table — and, since v1.2, that each county’s four tier shares sum to exactly 100.0 and that every occupation’s published tier matches its score recomputed at full precision.

One further check sits outside the automated gates. Before the site was built, a separately-written program — an AI agent working from the method description alone, without access to the pipeline code — independently recomputed the scores for three counties chosen for range (Denver, large and clean; Pueblo, mid-sized; San Juan, small and heavily suppressed). All three matched the pipeline at the published precision (two decimals) on both exposure score and high-exposure share, and the independent computation derived the same 50.0 threshold — the single cut then in use; the v1.2 tier boundaries are fixed values and need no derivation. This does not prove the method is right; it does establish that the published numbers follow from the stated method and inputs, rather than from undocumented code behavior.

Limitations

In rough order of importance:

1. National staffing patterns are assumed to apply locally. This is the core assumption of the whole method. A county’s hospitals are assumed to employ the national hospital occupation mix; its farms, the national farm mix. Real local deviations — a county whose manufacturing is unusually automated, or whose software firms are unusually sales-heavy — are invisible to us. The OEWS reconciliation bounds the aggregate damage at the state level (mean error 0.57 points by occupation group) but cannot bound it for any single county.
2. The staffing patterns are partly pre-2024. OEWS estimates pool six semiannual collection panels gathered over roughly three years — for the May 2025 file, approximately November 2022 through May 2025 — under model-based estimation. The staffing matrix therefore partially reflects the occupation mix of 2022–2024 workplaces, and any AI-driven change in staffing since then is smoothed away. The same pooling affects the published Colorado distribution used as the reconciliation target, so the reconciliation cannot detect it either.
3. “Employment” means UI-covered employment. QCEW counts jobs covered by unemployment insurance. It excludes most self-employed workers, most farm proprietors, and some others. In counties with high self-employment or proprietor farming, the atlas describes a smaller share of actual work than elsewhere. No figure on this site should be read as “all jobs.”
4. Jobs, not people. QCEW counts jobs; a person holding two covered jobs is counted twice. Shares of employment on this site are shares of jobs, not shares of workers.
5. Imputed is not observed. In 17 counties, more than half the underlying industry employment was suppressed and imputed from state-level mix. Those counties’ scores are consistent with the method, but the method is doing most of the work. The per-county imputation share is published precisely so readers can discount accordingly.
6. The scores are a 2023 snapshot of AI capability. Eloundou et al.’s annotations reflect what LLMs could plausibly do as of 2023. Capabilities have moved since and will keep moving; an occupation scored low because its tasks seemed out of reach may not stay low. Updating employment data alone does not update these capability ratings; a later edition would need a separately documented score refresh.
7. Exposure ≠ adoption ≠ displacement. Stated above, repeated here because it is the most common misreading: nothing in this atlas measures whether AI is being used in a county, or whether anyone has lost work to it.
8. Within-occupation variation is ignored. Every accountant in Colorado gets the same score. In reality, exposure varies with seniority, specialty, and firm — variation the occupation-level framework cannot see.
9. Known critiques of the exposure measures apply. The Eloundou framework treats an occupation as a bundle of independent tasks; if AI absorbs some tasks, the framework says nothing about how the remaining ones recombine into jobs. Its GPT-4-rated columns face a rater-circularity critique (a model grading its own reach) — we use the human-annotated columns, which dodge that specific objection but still reflect 2023 annotator judgment. The AIOE measure, used only as a robustness check, is ability-based rather than task-based and predates the current generation of systems.
10. The payroll-weighted figures use national wages. The share-of-payroll calculations weight occupations by national OEWS mean wages, not county wages. They are best read as “what if this county paid national average wages,” which is the standard limitation of this construction.

FAQ

Is my job going away? This atlas cannot tell you that, and does not try. A high score for your occupation means many of its tasks overlap with AI capabilities assessed in 2023. Historically, task overlap has meant the job changes — some tasks get faster, some get automated, new ones appear — more often than it has meant the job disappears. In one week of consumer conversations, model-classified usage (the observed-usage data above) leaned slightly more toward augmenting work than automating it — thin evidence, though that is the direction it points. But nobody can responsibly forecast employment for a specific occupation from this data, including us.

Doesn’t AI already do medical transcription? Why is it in the lowest tier here? Medical Transcriptionists score 23.2 — the raw Eloundou human-β rating (0.232, exactly 13/56) for the single O*NET occupation feeding this code, with no aggregation or imputation involved. The score measures prospective task overlap with LLM capability under that framework’s rubric; it is not a measure of the occupation’s automation history, and BLS employment data record large declines in this occupation from speech-recognition tools that predate the LLM era. Popular coverage tends to run those two stories together. The tier language on this site describes only the first.

Why is my county’s page marked “heavily imputed”? Because QCEW suppressed more than half of your county’s industry-level employment detail to protect employer confidentiality, and we filled the gap using state-level industry mix. The badge is there so you weight the number accordingly. The exact imputed share is printed on the page — the badge just makes it hard to miss.

A county with 1,100 jobs ranks 6th in the state. Should I believe that? Cautiously. Crowley County ranks 6th with 54.8 percent of its employment imputed, and a further point of fragility: 37.0 percent of its employment is fallback-routed — mostly public administration carried on the state/local government staffing blend — so its rank rides substantially on that blend being right. Several other small counties sit high in the ranking with similar flags. Small counties have small industry samples, heavy suppression, and single employers that can dominate the mix, so their point estimates are the least reliable in the atlas. That is exactly what the data-quality badge and the published imputation share are for. The rankings of the large-county pattern — Front Range metro counties high, agricultural counties low — rest on much firmer data. Since v1.1, counties this heavily imputed — Crowley among them — carry no ordinal in county-page prose or link previews; the band and the badge lead, and the rank stays in the table and the data files (see How counties are presented).

Why don’t your numbers match [national study]? Usually one of four reasons: a different exposure measure (we use Eloundou human-β; others use AIOE, or GPT-4-rated columns, or proprietary indexes), a different cut for the headline share (ours is the ≥45 high+substantial cut; see the exposure-measure section), a different employment base (we use UI-covered QCEW employment; ACS-based studies include the self-employed), or a different data year. The method section above gives enough detail to reconcile most gaps. If a discrepancy survives that reconciliation, we would like to hear about it.

Why can’t I see which specific occupations are exposed in my county? Because we don’t believe the imputation supports that claim, so we don’t publish it. County occupation detail stops at the 22 SOC major groups. Detailed-occupation county estimates produced by this method would carry false precision, and no downstream file of ours contains them. The occupation pages added in v1.1 do not change this: they report statewide Colorado figures from the published OEWS state file, not county estimates.

Does this include self-employed people and farmers? Mostly not. QCEW covers employment subject to unemployment insurance, which excludes most self-employment and most farm proprietors. See limitation 3.

Can I use these numbers? Yes. The data is licensed CC BY 4.0 — use it for anything, including commercially, with attribution. The full dataset, per-county files, and auto-generated data notes are on the Data page, and the citation format is below. The pipeline code is open source.

How will the 2027 edition differ? Same method, re-run on 2026 employment data and whatever exposure-score revisions the literature publishes. The new artifact will be the first year-over-year comparison; this 2026 edition is a single-year baseline. County URLs will not change; the 2026 data files remain permanently downloadable at /data/2026/.

Changelog & citation

September 2026 — wording and release housekeeping. The headline now identifies the counts as estimates, dates the capability ratings to 2023, and states that exposure is not a job-loss forecast. The contact address and ACS integration status are current. No scores, thresholds, employment figures, or dataset files changed.

v1.2 — tier revision (August 2026). No underlying value changed: employment, exposure scores, and imputation statistics are identical to v1.0. What changed: the binary ≥50 high-exposure cut is replaced by four tiers (cuts at 60, 45, and 30, inclusive lower bounds) with the headline moved to the high+substantial share (score 45 or above) — 27.4 percent statewide against 23.1 under the old cut, which remains published for every county as the secondary measure; county pages report the four-tier employment distribution; the bands, the median-county anchor, and the national anchor are recut on the new headline; the occupation files add tier labels and a flag for the 55 verified exact-zero scores; and the threshold section of this page was rewritten (see the exposure measure) to state the cut choice and show the alternative. Dataset version 1.2.

v1.1 — presentation revision (August 2026). No underlying value changed: employment, exposure scores, shares, ranks, and imputation statistics are identical to v1.0. What changed: county pages lead with the share of jobs (fraction and count) rather than the 0–100 mean score; five share-quintile bands with labels anchored to the typical county; ordinal ranks withheld from page prose and link previews for counties over 30 percent imputed; a same-method national anchor (21.7 percent); peer-county comparisons; a page per detailed occupation under /occupation/; illustrative occupation names under the county group bars; the “where AI lands last” selection corrected to groups strictly below the county’s own mean (the old lower-half rule could list a group scoring above the county mean under a “least exposed” heading); and, on this site, new methodology sections (presentation, occupation pages, what this can and can’t tell you) plus a separate, clearly labeled interpretation section rendered from site/content/interpretation.md. Dataset version 1.1.

v1.0 — 2026 edition. First publication. Employment data year 2025; Eloundou human-β exposure scores; 64 counties. Subsequent corrections, if any, will be listed here with dates and diffs — the underlying data files are versioned and the pipeline is deterministic, so any change to a published number will be visible and explained.

Requested citation:

Colorado AI Exposure Atlas, 2026 edition. Christopher Martin. https://coloradoaiexposureatlas.com/

Data: CC BY 4.0. Pipeline code: open source, MIT license.