Does AI exposure reach the dull, dirty and dangerous work it promises to relieve? A 754-occupation study says often not
An unreviewed US analysis linked 21 working-condition items to two AI-exposure measures. More exposed occupations generally had fewer selected burdens, and much of the LLM association weakened after physical demands and job characteristics were included.
The short version
- What happened
- The unreviewed study linked 21 O*NET working-condition items to two published AI-exposure measures across 754 US occupations representing 129.6 million jobs, or 92% of detailed employment in its source data.
- Who it affects
- Workers need task-specific evidence and a say in what changes; occupational exposure scores cannot identify their preferences.
- What remains unproven
- The paper is unreviewed, observational and not preregistered; secondary analyses are exploratory.
Keep in mind
The paper is unreviewed, observational and not preregistered; secondary analyses are exploratory.
Its 754 occupation averages cannot represent every workplace or worker, and the component datasets cover different years.
Exposure measures do not observe deployment, worker preferences, job loss, wellbeing or intervention outcomes.
The index omits many psychosocial and job-quality dimensions and does not directly measure boredom or desirability.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: current AI exposure is concentrated away from many measured burdens
A new unreviewed analysis finds that US occupations rated as more exposed to large language models and other AI generally score lower on a selected set of hazardous environments, adverse physical conditions, repetitive work and restricted discretion. That pattern challenges a familiar sales pitch. Automation is often justified as a way to relieve dull, dirty and dangerous work, yet the occupations most technically exposed to contemporary AI tend to be cognitive jobs with fewer of those measured conditions. The study does not prove that AI cannot help a construction worker, cleaner or nursing assistant. It shows that exposure scores alone are a poor map of where those burdens sit.
The comparison covers 754 occupations with usable employment data, representing 129,550,700 jobs and 92.0% of detailed national employment in the May 2021 Occupational Employment and Wage Statistics data. The author combined 21 items from the August 2026 release of O*NET into three domains. Dangerous included hazards and protective equipment; Dirty included outdoor work, noise, temperature, contaminants, cramped positions and vibration; Dull combined repetition, machine pacing and reversed measures of decision freedom. Higher values meant more of those selected conditions, not that every worker experiences or dislikes them in the same way.
Two exposure measures supplied different views of technical potential. The LLM measure came from the GPT-4-rated beta series in the published 'GPTs are GPTs' dataset and asks whether language models, alone or with software, could reduce task time. AI Occupational Exposure, or AIOE, links AI applications to abilities required in an occupation. Neither observes whether a tool was bought, used, accepted or rejected. Both partly use O*NET information, so agreement between them is not fully independent evidence. The paper is best read as an occupational measurement study, not as a forecast of job displacement or worker benefit.[1]
What the coefficients show—and what adjustment changes
The primary models compared occupations within 22 broad occupational groups and used both equal occupation weights and employment weights. A one-standard-deviation increase in LLM exposure was associated with a 0.350-standard-deviation lower composite condition score when occupations were weighted equally and a 0.285-standard-deviation lower score when weighted by employment. AIOE coefficients were larger in magnitude at minus 0.528 and minus 0.425. All 16 primary domain-by-exposure-by-weighting confidence intervals lay below zero after Holm adjustment; the largest adjusted p-value was 0.007. Those estimates establish a consistent descriptive direction within the chosen specifications.
The magnitude changed substantially when the analysis held the sample at 736 occupations and added familiar differences between jobs. With equal occupation weights, the LLM coefficient moved from minus 0.343 in the baseline to minus 0.093 after physical-demand adjustment, then minus 0.095 after also including job-zone indicators and log median wage. Its 95% interval ran from minus 0.214 to 0.023. With employment weights, the fully adjusted coefficient was minus 0.147, with an interval from minus 0.316 to 0.022. Both include no association. The adjusted AIOE intervals remained below zero: minus 0.186 [minus 0.320, minus 0.052] and minus 0.264 [minus 0.468, minus 0.060].
Adjustment is not a causal explanation. Physical demands overlap with the very burdens being summarized, while education, training, experience and wages do not capture every difference in task content or workplace power. The exercise asks how much unique association exposure adds after those characteristics enter the same model. The partial R-squared values were small: 1.1% for LLM exposure and 2.4% for AIOE with equal occupation weights, rising to 3.4% and 7.6% with employment weights. Those figures qualify the sweeping claim that an exposure index identifies where AI can relieve undesirable work.[1]
The distinction that matters
Why exposure is not the same as relief
The study links occupational averages; every step between technical potential and a better job remains unobserved.
Conditions
Combine 21 O*NET items
A relative occupational burden score
Exposure
Link LLM and AIOE measures
Technical potential, not adoption
Adjustment
Add demands, preparation and wages
Much smaller LLM association
Relief
Ask workers and test a deployment
Evidence the study does not provide
A single 'dull' score hides repetition and control
The paper separates repetition and machine pacing from restricted discretion because they are not the same experience. Under equal occupation weights, the LLM coefficient was minus 0.102 for repetition, with a 95% interval from minus 0.233 to 0.028, but minus 0.334 for restricted discretion, with an interval from minus 0.452 to minus 0.216. With employment weights, the repetition interval reached from minus 0.242 to 0.001, while the restricted-discretion coefficient was minus 0.458 [minus 0.759, minus 0.156]. The author did not formally test the difference between those coefficients, so it should not be presented as a proved contrast.
This distinction matters for design. A system may shorten data entry without giving a worker greater control over priorities, staffing or pace. Conversely, repetitive manual movement may be unaffected by a language model even if the occupation contains administrative tasks that score as exposed. Calling both situations dull blurs the task an intervention changes and the condition a person wants changed. The O*NET index also does not measure boredom, meaning, satisfaction or whether protective equipment is experienced as a burden. A high score means more of the coded conditions, not a universal ranking of bad jobs.
The occupational profiles make the same point. Construction labourers, heavy truck drivers, nursing assistants, restaurant cooks, landscaping workers and housekeeping cleaners all appeared in the above-median-burden, below-median-exposure group. Yet construction labourers scored highest in the Dirty domain, while housekeeping cleaners scored highest in the combined Dull domain and below the sample average in both physical domains. Across separate thresholds, 31.6% of employment fell above the condition median and at or below both exposure medians; a stricter tertile definition still covered 15.6%. Those are relative classifications, not thresholds of harm or feasibility.[1]
What employers, workers and researchers can responsibly take from it
For employers, the result argues against choosing automation projects from an occupation-level AI score. A credible proposal should name the task, the condition it changes, the people affected and the trade-offs. A dispatch assistant might alter planning work while leaving a driver's road risk, schedule pressure or physical demands untouched. A documentation tool might save a nurse time while staffing rules consume the saving. Evaluation should measure the claimed outcome—hazard exposure, decision latitude, workload or recovery time—and ask workers whether that change is valuable, rather than treating faster task completion as relief by definition.
For workers and representatives, exposure is not destiny. The indices represent technical potential averaged over an occupation, not within-job variation, bargaining power, deployment quality or management choices. Low LLM exposure also says little about robotics, mechanisation or algorithmic management. The source data exclude self-employment from employment weights, aggregate detailed specialties and combine datasets from different years. Confidence intervals use 22 occupational clusters and omit uncertainty from item selection, source ratings and crosswalks. The study was not preregistered, and its adjustment, decomposition and secondary analyses are exploratory.
The preprint discloses that generative AI assisted code development, literature discovery and manuscript drafting and revision; the author retains responsibility. It uses public aggregate data and collected no participants. The next evidential step is not a larger occupational correlation alone. It is a prospective study that begins with worker preferences, observes a specific implementation and measures whether conditions improve without shifting costs elsewhere. Replication with workplace-level data, other countries and technologies beyond language models would also change confidence. Until then, promises to remove dull, dirty and dangerous work should be tested task by task, with workers defining what relief means.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring automation: compare the finding and its evidence limits.
If robots can do much physical work, why are so few tasks economical to automate?
Anthropic researchers rate 7,594 physical job tasks and estimate that robots can perform 74% in some setting, yet are currently cheaper than people for only 0.3% of all work. The gap between technical exposure and economic adoption is the result that matters for workers.
Read the next reportExplore the full source and evidence record
What this means for people
- Workers need task-specific evidence and a say in what changes; occupational exposure scores cannot identify their preferences.
- Employers should measure whether an implementation reduces a named hazard or burden, not only whether it speeds up work.
- Unions and regulators can use the profiles to choose contrasting workplaces for research, but not as a priority ranking or proof of benefit.
Global context
The analysis uses US O*NET and labour data. Jobs, protections, informality, technology access and bargaining institutions differ internationally. Replication needs local occupational measures and direct worker evidence rather than a mechanical transfer of US coefficients.
What the evidence does not yet show
- The paper is unreviewed, observational and not preregistered; secondary analyses are exploratory.
- Its 754 occupation averages cannot represent every workplace or worker, and the component datasets cover different years.
- Exposure measures do not observe deployment, worker preferences, job loss, wellbeing or intervention outcomes.
- The index omits many psychosocial and job-quality dimensions and does not directly measure boredom or desirability.
What to watch next
- Peer review and independent replication.
- Workplace trials that predefine a condition and measure change.
- Worker preference data linked to specific tasks.
- Comparisons with robotics and algorithmic-management exposure.
Living evidence record
Impact record IAI-1YBRBCE
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Whether occupations rated as more exposed to language models and AI also contain the physical hazards, adverse environments, repetition and restricted discretion invoked by promises to automate dull, dirty and dangerous work
Topics: Future of work · Occupational exposure · Worker agency · Automation
Regions: United States · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.