Artificial Intelligence Exposure, Relational Work and Labour-Market Outcomes: A task-based analysis of 741 US occupations using O*NET tasks, LLM exposure scores and BLS projections
Translated title
Artificial Intelligence Exposure, Relational Work and Labour-Market Outcomes
Author
Ahmed, Ahnaf
Term
4. semester
Publication year
2026
Submitted on
2026-10-14
Abstract
Public discussion of generative artificial intelligence (GenAI) and robotics often focuses on which tasks machines can perform. This thesis adds another perspective: how much the value of a task depends on human presence, trust, and social interaction – in other words, its relational content. The study analyzes 16,176 task statements from O*NET, a U.S. database that describes the content of occupations. Each task was rated by a large language model (Gemini 3.1 Pro) on a 1–5 scale for (1) exposure to GenAI, (2) exposure to robotics, and (3) amount of relational content. Task scores were then combined with information on how frequent and important each task is, to produce average scores for 741 detailed occupation categories (six-digit SOC codes). These occupations were matched with Bureau of Labor Statistics (BLS) data on projected employment changes for 2024–2034 and median wages for 2024. The study is descriptive and correlational, using Spearman correlations, partial correlations (which statistically control for other variables), and k-means clustering (a method for grouping occupations with similar profiles). To check the reliability of the ratings, a random sample of 100 tasks was re-scored by a second model (GPT-5.5). Agreement between the models was high (quadratic-weighted κ = 0.71–0.72), although relational scores were more sensitive to which model was used. All three measures – GenAI exposure, robotics exposure, and relational content – are positively associated with skill requirements. In occupations with higher Job Zones (indicating more education and experience), GenAI exposure and relational content tend to be higher, while robotics exposure tends to be lower. When other factors are controlled, relational content shows the strongest positive correlation with projected employment growth (ρ = 0.31). Within Job Zones, higher GenAI exposure is weakly negatively related to projected growth (partial r = -0.11) but positively related to wages (r = 0.14). Higher relational content is positively related to employment growth (r = 0.15) but not to wage levels. The analysis identifies four exposure profiles across occupations. Among occupations that are strongly exposed to GenAI, those with more relational content have better projected employment outcomes in mid-skill Job Zones 2–3, but not in higher-skill Job Zones 4–5. Overall, the findings suggest that relational content may act as a partial, skill-dependent buffer against technological displacement. This protective effect appears more clearly in employment prospects than in wages. Keywords: generative AI, robotics, relational work, task-based approach, O*NET, BLS, Job Zone, occupational employment projections, k-means clustering
Debatten om generativ kunstig intelligens (GenAI) og robotteknologi handler ofte om, hvilke konkrete opgaver teknologi kan overtage. Denne afhandling tilføjer en ny dimension: hvor vigtig menneskelig tilstedeværelse, tillid og social kontakt er for værdien af en opgave – det vil sige opgavens relationelle indhold. Studiet analyserer 16.176 opgavebeskrivelser fra O*NET, en amerikansk database over jobindhold. Hver opgave blev vurderet af en stor sprogmodel (Gemini 3.1 Pro) på en skala fra 1 til 5 for: (1) udsættelse for GenAI, (2) udsættelse for robotteknologi og (3) mængden af relationelt indhold. Disse opgavescorer blev kombineret med oplysninger om, hvor ofte og hvor vigtige opgaverne er i jobbet, så der kunne udregnes gennemsnitsscorer for 741 detaljerede jobkategorier (SOC-occupation codes). Jobkategorierne blev derefter koblet med data fra Bureau of Labor Statistics (BLS) om forventede ændringer i beskæftigelse 2024-2034 og medianløn i 2024. Undersøgelsen er beskrivende og korrelativ og anvender både Spearman-korrelationer, partialkorrelationer (hvor andre variable statistisk kontrolleres) og k-means clustering (en metode til at gruppere job efter lignende karakteristika). For at teste om vurderingerne var stabile blev et tilfældigt udvalg på 100 opgaver vurderet igen af en anden sprogmodel (GPT-5.5). Overensstemmelsen mellem modellerne var høj (quadratic-weighted κ = 0,71-0,72), men den relationelle score var mere følsom over for, hvilken model der blev brugt. Alle tre mål – GenAI-eksponering, roboteksponering og relationelt indhold – hænger positivt sammen med jobkrav og færdighedsniveau. I job med højere Job Zone (et mål for, hvor meget uddannelse og erfaring der kræves) stiger GenAI-eksponering og relationelt indhold, mens roboteksponering falder. Når der kontrolleres for andre faktorer, har relationelt indhold den stærkeste positive sammenhæng med forventet jobvækst (ρ = 0,31). Inden for de enkelte Job Zones gælder, at højere GenAI-eksponering er svagt negativt forbundet med forventet vækst (partial r = -0,11), men positivt forbundet med løn (r = 0,14). Mere relationelt indhold er positivt forbundet med vækst (r = 0,15), men ikke med lønniveauet. Analysen identificerede fire typer eksponeringsprofiler for job. Blandt job, der i høj grad er udsat for GenAI, havde de mere relationelle job bedre forventede beskæftigelsesudsigter i de mellemste færdighedszoner (Job Zone 2-3), men ikke i de høje færdighedszoner (Job Zone 4-5). Samlet peger resultaterne på, at relationelt indhold kan fungere som en delvis, færdighedsafhængig buffer mod teknologisk fortrængning. Denne beskyttende effekt ses primært i beskæftigelsen og ikke i lønudviklingen. Nøgleord: generativ AI, robotteknologi, relationelt arbejde, opgavebaseret tilgang, O*NET, BLS, Job Zone, beskæftigelsesfremskrivninger, k-means clustering
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords