750 Million Job Ads Show AI Erases Job Titles Faster Than Tasks

Two slopes, measured on the same pile of 752.6 million Chinese job advertisements, disagree about what AI is doing to work. [Fact] Occupations built mostly of AI-exposed tasks lost posting share between 2022 and 2026 at a slope of -1.33. [Fact] The broad task categories those occupations are made of declined far more slowly, at -0.83. If AI were hollowing jobs out from the inside — deleting the automatable tasks and leaving a leaner job behind — the second number should be the steeper one.

It is not. And that inversion is the sharpest finding in "The Pulse Beneath the Job Title" (arXiv:2608.26924), a preprint posted on August 27, 2026 by Qin Chen, Ying Fang, Xiangyu Wang, and Leo Yang Yang. [Fact] The authors compress it into one sentence: the tasks outlive the jobs that carried them.

A task observatory built on roughly $460 of judgment calls

The raw material is every posting from five major Chinese recruitment platforms between January 2022 and June 2026: 752.6 million ads, growing from 30 million in 2022 to 287 million in 2025, and deduplicating to 116 million unique job descriptions. [Fact] From a stratified sample of 2.31 million descriptions — about 2% of the unique pool — an LLM extracted atomic phrases, one action, credential, or condition each, yielding 2.9 million raw requirement phrases and 7.3 million task phrases. [Fact]

The clustering step is where the engineering gets interesting. Embedded phrases were grouped into 262,144 initial clusters, and an LLM then adjudicated roughly 520,000 boundary pairs — deciding, for instance, that "maintain client relationships" should absorb 18 clusters and 797 surface variants — at a total adjudication cost of about ¥3,300, which converts to roughly US$460. [Fact for the yuan figure; the dollar conversion is our arithmetic] [Estimate] The output is a catalog of 20,721 requirements sorted into 16 classes and 181 subclasses, plus 44,479 tasks organized as 32 business domains crossed with 26 action modes. [Fact]

For scale: O*NET, the U.S. reference directory our own site builds on, carries 19,281 tasks and reports one national average per occupation, refreshed over multi-year survey rounds. [Fact] This system re-reads the entire posting flow monthly, and fewer than 0.005% of postings match nothing in either catalog. [Fact] The exposure scores it produces are not free-floating either: at the occupation level they correlate 0.73 to 0.79 with published GPT-exposure and AI Occupational Exposure indices. [Fact]

The adjustment runs through job titles, not job content

Aggregate AI exposure of posted labor demand in this data has been declining since late 2023. [Fact] The question is the channel. A shift-share decomposition splits the decline into two possible routes: employers posting fewer high-exposure occupations, or employers keeping the occupations but stripping the exposed tasks out of them. The paper's answer is close to unanimous — the gradient "runs almost entirely through the posting-share channel," while within-occupation task composition stays flat, with "no systematic stripping of exposed content." [Fact]

Put the two slopes side by side and the asymmetry is stark. The occupation-level gradient of -1.33 is about 1.6 times steeper than the task-level gradient of -0.83 — and in a robustness check counting each unique description once, the occupation slope steepens to -1.76, stretching that ratio to roughly 2.1 times. [Fact for the slopes; the ratios are our own calculation, which the paper does not report] [Estimate]

Employers, in other words, are not rewriting jobs. They are declining to post them.

There is an obvious objection, and the authors raise it themselves: the pattern is descriptive, "as consistent with cyclical stories as with technological ones." [Fact] China's 2022-2026 window was not a calm one, and the dataset leans hard toward exactly the sectors where AI exposure and hiring cycles tangle — information and software makes up 13% of postings against 1.7% of census employment. [Fact] A cooling tech cycle would produce some of this signature without any AI at all. The monthly cadence is what makes the question answerable later; it does not answer it yet.

One job title, a 31-point spread in AI exposure

The second finding attacks the idea of "an occupation" itself. Take production workers. The task bundle advertised in the lowest wage band carries an LLM exposure of about 17%; in the highest band, about 48%. [Fact] That is a 31-percentage-point spread inside a single job title — our subtraction, not the paper's — because high-wage production postings emphasize quality documentation and process coordination rather than assembly. [Estimate] Note the direction: the better-paid version of this job is the more AI-exposed one, an inversion of the comfortable assumption that moving up always means moving to safety.

Accountants show the same internal stretch. As posted wages rise from under ¥5,000 to ¥30,000-50,000 a month, core bookkeeping falls from 68% to 49% of the task mix — a 19-point recomposition toward compliance, data analysis, and administration. [Fact] Certification demands climb from 28% of postings to 52% at ¥12,000-20,000, and there they stop; above that band, the entry-level certificate collapses from 14.5% to 4.7% of postings while CPA demand more than triples, attitude-type requirements shrink from 22.4% to 14.2% of mentions, and experience rises from 8.8% to 15.2%. [Fact]

And then there is the counterexample that keeps the finding honest: Java developers carry nearly the same task bundle at every experience level, with only 3-5% of postings asking for any certificate. [Fact] Some job titles are honest averages. Others are camouflage.

The uncomfortable part for one-number-per-job sites — including this one

This paper is a direct methodological challenge to any site that reports a single automation-risk figure per occupation. That includes ours. Our pages for accountants and software developers summarize O*NET-derived, occupation-level evidence — useful as a starting point, but exactly the kind of one-average-per-title measure this data shows can conceal a 17%-to-48% internal range. [Claim] If within-title variation is this wide in Chinese posting data, the honest reading of any single occupation-level number, ours included, is "the center of a wide distribution," not "your personal risk."

What the data cannot say

A posting is a vacancy advertisement, not a hire; the system measures expressed demand, not employment outcomes. [Fact] The data over-represents urban white-collar hiring — tier-1 and tier-2 cities supply 54.8% of postings but only 20% of census employment, sales and service occupations are at 49.6% of postings versus 33.9% of employment, agriculture (20.5% of Chinese employment) is essentially absent, and construction is under-represented at 2% of postings versus 11.3% of employment. [Fact] The decomposition is descriptive, not causal. And this is a version-1 preprint, not yet peer-reviewed — whether the "titles first, tasks later" pattern transfers to labor markets outside China is an open question this single-country dataset cannot settle. [Claim]

What workers should watch

Three practical readings. First, if adjustment runs through posting share, AI pressure hits job seekers at the hiring gate before it hits incumbents in their task lists — so watch vacancy volumes in your field, not just whether your own workload feels automatable. Second, within-title spread cuts both ways: for accountants, the higher bands shift toward compliance and analysis, while for production workers the higher bands are more exposed, so check what your next wage band actually asks for rather than assuming promotion equals protection. Third, credentials have a ceiling — in the accountant data, certificate demand plateaus at the middle band and the top bands trade credentials for experience and judgment. Past a point, another certificate is not the lever.

AI-assisted analysis: this article was drafted with AI assistance from the arXiv preprint cited below and reviewed before publication. Figures are quoted from the paper; calculations marked as ours are derived from those figures.

Sources

• Qin Chen, Ying Fang, Xiangyu Wang, Leo Yang Yang (2026). "The Pulse Beneath the Job Title: Monthly Readings of Requirements and Tasks from 750 Million Chinese Job Ads." arXiv:2608.26924, submitted August 27, 2026. Licensed CC BY 4.0. https://arxiv.org/abs/2608.26924