One in Five Job Ads in One Hiring Network Mentions AI. Most of That Is the Fine Print.

There are three ways to ask a corpus of job postings how much employers want AI, and they give three different answers. You can count the ads that mention AI at all; the ads that require an AI skill; or the ads for a job that is an AI job. On this corpus, this spring, those counts were one in five, one in fourteen, and one in 140. The distance between them is the story — because most of what looks like an AI hiring boom is an employer telling you it has heard of AI, not asking you to build with it. This is an account of that gap, measured across one applicant-tracking network over three months, and of what a single-source record can and cannot certify about who is buying AI work.

What this corpus can see

The scope governs every reading that follows. In April, May, and June 2026 — the primary window every claim here rests on — the corpus holds about 2.87 million distinct U.S. vacancy-locations, the unit this publication counts: a single opening at a single place, deduplicated against the same opening re-posted or re-scraped. It is not a survey of employers and not a census of vacancies; it is a record of what employers chose to post online.

One fact dominates, as it did in our first study. Every month, 93–99% of these postings come from a single source: DirectEmployers, an applicant-tracking network used by large U.S. employers. So the corpus does not observe "the labor market." It observes direct-employer postings from large U.S. employers, and every claim below is scoped that way. A second fact fixes which months can carry a claim: February is 87% one ingest day and March is a 2.19-million-row burn-in backfill, so both are drawn but greyed, never load-bearing. Claims rest on April–June; July is partial.

How "AI" is defined matters, because the whole finding turns on it. We flag every posting three independent ways, as regex tiers, disclosed rather than hidden. Core (tier 1): the job itself is an AI role — ML engineer, applied scientist, prompt engineer. Skill (tier 2): the ad requires an AI competency — machine learning, LLMs, PyTorch, RAG, a named model. Mentioned (tier 3): any AI reference at all, including one line of boilerplate. The tiers are independent booleans; every number states its tier. One seam is worth naming up front: about 22% of tier-2 hits fire on soft "experience with / exposure to / understanding of AI" phrasing anchored to a bare AI acronym, not a named technique or tool — so the tier-2/tier-3 boundary is fuzzier than "concrete competency" implies. Restrict tier-2 to a named skill and the 7.4% below falls to about 5.7%; we report the broader definition and flag the soft share rather than pick one.

The funnel

Here are the three tiers as a share of all new postings, by month, on the primary window:

Shares = each tier's VL ÷ total_vl.

Pooled across the primary window, 20.7% of ads mention AI, 7.4% require an AI skill, and 0.71% are an AI role — one in 4.8, one in 13.5, one in 141. Read the tiers against each other and the point lands: of the ads that mention AI, only about 36% ask for any AI skill and only about 3% are an actual AI job. Nearly two in three ads that name AI do not want you to know a thing about it.

The two upper tiers rose across the quarter; the bottom one did not. Mentions climbed 6.3 points April to June and AI-skill demand 1.9 points, but the AI-role share sat flat at roughly 0.71% the whole way. A batch-day jackknife — dropping each of the ten highest-volume ingest days one at a time — moves the June skill share by at most 0.59 point and the June mention share by 0.37 point, inside the trends, not explaining them.

Part of the rise is the corpus changing shape, not employers changing behavior, and an occupation shift-share separates the two. Of the April-to-June AI-skill-share rise, roughly a third (36%) is occupation-mix drift — more computer-and-technical postings entering the corpus — rather than employers within a given occupation asking for AI more often; about a fifth (20%) of the mention rise is likewise composition (the decomposition runs on SOC-coded postings, where the raw skill rise is a shade steeper than the corpus-wide 1.9 points). The within-occupation rise is real but smaller than the headline point-change — only about two-thirds of the skill rise is employers within a job asking for AI more. The AI-role tier is flat on both cuts (a +0.02-point change April to June, either raw or mix-adjusted). So what is rising is talk and tooling — and, in the technical occupations, some genuine within-occupation skill demand — but not AI headcount.

Most of it is the fine print

Why does "mentions AI" run three times "requires AI"? Because much of the mention is a single line an employer stamps onto every ad, regardless of the job. Rank employers by how often their ads mention AI and a distinct class appears: firms that put an AI reference in 90% or more of everything they post. There are 163 of them above a 100-posting floor, and together they account for 48% of every AI-mentioning ad in the corpus — a share that runs from 43% under a strict 95% definition to 53% under a looser 70% one.

That 48% is not all marketing boilerplate, and the number cuts both ways. About one in seven of those mentions (14%) comes from genuinely AI-intensive employers — Meta, Anthropic, Anduril, Coinbase, NVIDIA, OpenAI — whose ads mention AI in nearly every posting because the role really is AI, not because a stock paragraph rides along. Strip those firms out and firm-wide marketing boilerplate still accounts for roughly 41% of AI-mentioning ads. In the other direction the 41% is itself conservative, because the firm-wide test catches only employers whose every ad carries the line, not the incidental boilerplate buried inside more varied companies. Call it two-fifths to a half, and do not read the whole 163-firm class as fine print: some of it is the job.

The examples make it concrete. Oracle posted 70,866 ads; 99.8% mention AI — Oracle brands its platform as AI-powered, so the word rides along on every listing, from database roles to sales. Intermountain Health mentions AI in 13,347 of its 13,364 ads (99.9%) and asks for an actual AI skill in fewer than one in a hundred. Robert Half Accountemps, an accounting-temp agency, mentions AI in 6,296 of 6,317 postings — and names a specific model in a literal handful. That is the fine print: an "AI" that appears in accounting-clerk and hospital-support ads because it lives in a company's stock paragraph, not its job.

The boilerplate share falls as the tier gets stricter — 48% of mentions come from these firms, but 36% of skill requirements and 28% of AI roles. The deeper past the marketing line, the more the signal is real. Strip the boilerplate employers out and the organic mention rate still rises across the quarter (9% to 14%) — so this is not only boilerplate — but the level is inflated by exactly the language the Budget Lab has called "AI-washing," and this corpus shows its mechanism.

Which jobs actually want the skill

Where AI-skill demand is real, it is steep and narrow. At the occupation grain the crosswalk measures well — SOC minor groups, ~84% precision — computer and mathematical occupations tower over everything: mathematical-science occupations at 44% AI-skill, computer occupations at 30%, against food-and-beverage serving at 0.05% and K–12 teachers at 0.01%. At a finer six-digit grain the split looks sharper still — Computer and Information Research Scientists appear to require an AI skill in 97% of ads and to be an AI role in 63%, Data Scientists 54% and 38%, against Accountants at 3.3% and bookkeeping clerks at 2.5% — but these six-digit rates are illustrative colour only, not measured penetration. The soc6 match tier they rest on blends an exact-string component the crosswalk audit clears at just ~58% strict precision, and the title-to-SOC assignment for research-scientist codes is partly self-fulfilling: a title bank that maps "applied scientist" into the research-scientist code will then measure that the code requires AI. Read 97/63/54/38 as direction, never as a penetration rate. The safe reading is the minor-group one: AI-skill demand is not spreading evenly; it concentrates in the occupations that were already technical — the gradient the exposure literature (Eloundou et al.; Felten–Raj–Seamans) predicts for generative rather than routine-task automation.

The tools employers name

When an ad names a specific model, a few dominate. Across the primary window, Copilot leads at 0.56% of all ads, then Claude (0.50%), ChatGPT (0.36%), and Gemini (0.31%). Read these as primary-window levels. Across the three clean months Copilot, Claude and ChatGPT all climb steadily — ChatGPT roughly doubles, from 0.20% of ads in April to 0.44% in June — while Gemini eases back below its spring peak; the greyed February–March months are drawn but never load-bearing. Treat the ranking as which assistants employers now name, not their market share.

Who is buying, and where

AI roles concentrate in a handful of employers. The five largest AI-role posters — Deloitte, EY, Capital One, PwC, and Amazon — account for 32% of every AI-role posting, and the top ten for 45%. The leaders are professional-services and cloud firms staffing AI for other companies as much as building it themselves; read this as posting behavior, not headcount. By company size, AI-skill demand is a large- and mid-employer phenomenon — 8.1% of large-employer ads and 10.0% of medium ones require an AI skill, against 3.9% of small ones (an R5 preview, not its finding). Do not read the medium-over-large ordering as real: it is an occupation-mix artifact that disappears once occupation is held constant — standardized to a common mix the two bands tie at about 8% (see R5). The only size gap that survives standardization is the smallest employers' shortfall.

Geographically the density is where you would guess, and steeper. Among computer-and-mathematical postings the AI-skill share runs 57% in San Jose, 51% in San Francisco, and 49% in Seattle, falling to the high-30s and below across the other large metros. These metro shares are computed at the SOC-15 minor-group grain and are not adjusted for finer sub-occupation mix within computing — San Jose skews toward research-scientist titles, so part of its lead is which computing jobs it posts, not a higher AI rate for the same job. Across all occupations San Jose still leads at 28%; a few small state-capital metros post high shares too, but on employer-concentrated volume — single-employer shares near the 50% suppression cap — and they reflect a narrow government-and-tech posting base rather than a broad local AI labor market, so they are flagged rather than ranked. (Their raw posting counts are not thin: several capitals carry more volume than San Jose; concentration, not size, is why they are down-weighted.)

Methodology

Every number is drawn from one corpus snapshot dated July 12, 2026, and reproduces from the six-table query cube built from it. The method — flow rather than stock; shares, never raw counts; the April–June primary window with February and March greyed; the mandatory batch-day jackknife; no absolute comparison to government data; no causal language; the DirectEmployers scope limit; and the three-tier regex AI definition disclosed in full — is set out on the methodology page, against which this article is written to be checked. The AI tiers are a full-text re-stream keyed to each posting, with documented false-positive guards (bare "AI" versus the EEO race code and Adobe's .ai; "LLM" versus the law degree; "machine learning" versus sewing machines).

Limitations

• Scope is one applicant-tracking network. 93–99% of postings are DirectEmployers. Every claim is scoped to direct-employer postings from large U.S. employers, not "the labor market," and the literature (Cammeraat & Squicciarini) documents that online postings over-represent computer, business, and professional occupations — the very ones this finding concentrates in.
• "AI" is a regex definition, in three disclosed tiers. The tiers are independent booleans (core ≈ 0.70%, skill ≈ 7.1%, mentioned ≈ 18.0% across all drawn months); every number states its tier. A different vocabulary would move the levels. One soft edge: about 22% of tier-2 ("AI skill required") hits fire on "experience with / exposure to / understanding of AI" phrasing anchored to a bare AI acronym, not a named technique or tool, which blurs the tier-2/tier-3 boundary; restrict tier-2 to a named competency and the 7.4% primary-window share falls to about 5.7%. We ship the broader definition and flag the soft share rather than pick one.
• The boilerplate split is a threshold, and it cuts both ways. Classifying an employer as "firm-wide boilerplate" at a ≥90% mention rate is a choice; we report the 43–53% sensitivity band across 70–95% thresholds. The 48% is contaminated upward by genuinely AI-intensive firms — about one in seven of the bucket's mentions (Meta, Anthropic, Anduril, Coinbase, NVIDIA, OpenAI) whose ads mention AI because the role is AI; net of them, firm-wide marketing boilerplate is roughly 41%. It is contaminated downward because the firm-wide test misses incidental boilerplate inside more varied employers. Two-fifths to a half is the honest range, not a one-sided lower bound.
• Trends rest on three months. February and March are greyed (batch-fragile and burn-in); July is partial. The rising skill and mention shares survive the jackknife, but three post-burn-in data points are a short series. A shift-share decomposition shows part of the rise is the corpus tilting toward technical occupations, not employers changing behavior: about a third of the AI-skill rise and a fifth of the mention rise is occupation-mix drift (gshiftshare_trend.csv), and a rising mention share further reflects more employers adopting AI language, which is not the same as more AI work.
• Occupation grain. About 74% of primary-window postings carry a SOC code; occupation shares are within covered postings. Minor-group claims sit at ~84% crosswalk precision and are the publishable grain. The six-digit contrasts (Research Scientists, Accountants) are directional colour only: the soc6 match tier blends an exact-string component the audit clears at just ~58% strict precision, and title-to-SOC assignment for research-scientist codes is partly self-fulfilling, so the 97/63/54/38 rates are vivid illustration, never measured penetration.
• Named-tool figures are primary-window levels. All four named assistants (Copilot, Claude, ChatGPT, Gemini) appear across the drawn months; the February–March values are greyed context, not load-bearing. Across the clean April–June window Copilot, Claude and ChatGPT rise and Gemini eases, but with three post-burn-in points these are reported as levels plus direction, not a fitted trend.
• Employer and size figures are posting behavior, not headcount, and the size band is a public-knowledge estimate (R5's subject), not registry data.