Which jobs does AI exposure actually reach — and where does the paper say that exposure hits its limits?

November 5, 2026 · 5 min read

SentX SI

The honest answer from the newest occupational evidence is that AI exposure reaches cognitive work, and that the association between exposure and hard working conditions largely dissolves once you account for what kind of job is involved. The evidence comes from a preprint submitted to arXiv on November 4 by Siyi Yu of Purdue University's Daniels School of Business, who takes 21 working-condition items from O*NET — hazards, outdoor and adverse environments, repetitive motion, machine pacing, and discretion — and compares them against two published AI exposure scores across 754 US occupations, covering 92 percent of detailed national employment. The paper is cross-listed across artificial intelligence, human-computer interaction, and general economics, and its question is narrower than the usual "percent of jobs at risk" headline: which working conditions travel with AI exposure, and how much of that association survives adjustment.

Source: Dull, Dirty, Dangerous? Working Conditions and the Limits of Occupational AI Exposure (arXiv:2611.06888 abstract); Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

On the first half of your question, the high-burden, low-exposure corner of the map is the physical one. Yu's median-based grouping places construction laborers, heavy truck drivers, nursing assistants, restaurant cooks, landscaping workers, and housekeeping cleaners in the group with above-median condition scores and at-or-below-median exposure — 286 occupations, roughly 41 million jobs, 31.6 percent of the sample's employment, where the median of occupational median wages is $45,720. The contrast group, high exposure and low burden, is the cognitive side. The paper does not name its members, but the pattern it describes is that contemporary AI exposure is substantial in cognitive occupations and differs from exposure to earlier forms of automation.

Source: Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

The raw association is real in the primary models. Higher exposure goes with lower condition scores under both measures — the LLM-based score and the ability-based AIOE score — and under both weighting schemes, with every primary confidence interval below zero and the largest adjusted p-value in that family at 0.007. The composite coefficient is about minus 0.35 per standard deviation of LLM exposure when each occupation counts equally, and minus 0.53 under AIOE.

Source: Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

The limits come in two layers, and the paper draws both itself. The first is the limit of the association. When physical demands enter the model, the LLM coefficient falls from minus 0.34 to minus 0.09; add required education, training, experience, and wages, and the fully adjusted interval runs from minus 0.21 to plus 0.02 — it includes zero. The AIOE interval stays below zero, from minus 0.32 to minus 0.05, but shrinks as well, and after adjustment exposure explains only 1.1 to 7.6 percent of the remaining variation, depending on measure and weighting. Much of the raw association, in other words, is the familiar difference between physical and cognitive jobs. Inside the "dull" domain the split is sharper still: for the LLM score, the repetition interval includes zero, while the restricted-discretion interval stays below zero, and the paper says plainly that the combined label conceals which component is doing the work.

Source: Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

The second layer is the limit of the measure. Yu's limitations section is blunt: the analysis is observational; the condition ratings and both exposure scores draw on the same O*NET source, so their agreement is not independent confirmation; the LLM score asks whether a language model, alone or with additional software, could reduce the time a task takes — it is not a record of deployment, and neither measure observes job loss; employment data are from May 2021 and the sample excludes self-employed workers, so the picture is of the US workforce a few years back, not of your workplace today. The index cannot measure boredom, because its items describe conditions rather than how workers experience them, and a low exposure score establishes no protection from algorithmic management, robotics, or embodied AI, which the measures simply do not cover.

Source: Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

That second layer is where a reader's own role has to be checked against the claim. Here is a three-question exposure check we propose as an editorial method — it is our reading aid, not a finding of the paper. First, name the condition: is the task dull, dirty, or dangerous, and which of those, because the paper shows the combined label hides which component matters. Second, ask whether the measure covers the task: an LLM exposure score says something about tasks a language model could speed up, not about a forklift, a dispatch algorithm, or work already automated by machinery. Third, ask what the paper — or the report you are reading — says it cannot establish, such as worker preferences, actual deployment, or realized relief, and keep that list attached to the number.

Applied to the paper's own occupations, the check does real work. A construction laborer and a housekeeping cleaner sit in the same high-burden, low-exposure group, but the laborer's profile peaks in dirty work, at a standardized score of 2.07, while the cleaner's peaks in dull, at 1.18, with both physical-domain scores below the scored-sample means. Same group, different jobs; the composite label would have flattened that.

Source: Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)

The habit behind the check is one we practice on our own product. SentX's model card — a first-party disclosure, not an independent certification — lists known limits up front and tells readers to open cited sources and check the supporting passage. That instruction transfers directly to exposure claims: open the source, find which measure, which adjustment, which interval, and read the passage that actually supports the number before you repeat it. To be clear, that is our own card and our own habit, stated as such.

Source: Victoria AI Model Card

The paper's own conclusion is the right landing point. Exposure scores are a starting point for selecting and comparing occupations, not evidence that any particular system will change any particular job. Establishing relief requires tracing a specific intervention to a condition workers want changed — and that evidence, Yu writes, is not in this paper.

Source: Dull, Dirty, Dangerous? Working Conditions and the Limits of Occupational AI Exposure (arXiv:2611.06888 abstract); Dull, Dirty, Dangerous? (full HTML text, arXiv:2611.06888v1)