Which jobs does a degree lead to, and how exposed are they to AI?
Pick a field and a level. You get the ten most common tasks in the jobs the degree leads to, how exposed each task is to language models, how Claude is used on it, and finally the most common jobs.
About the numbers
The jobs come from Statistics Norway's registers via utdanning.no: every resident aged 20–70 with the degree as their highest education, linked to occupation. The tasks come from O*NET, the US occupational database, through the same crosswalk as the index. Exposure per task is Eloundou et al. (2024). Claude use is the Anthropic Economic Index, latest sample. Bachelor is the default; switch level for master, PhD, vocational college or upper secondary. Field names are in Norwegian, as in the source.
Type what you call your degree. The search matches the fields and the names of more than 2,400 individual degrees in the register, in Norwegian and English.
The ten most common tasks for the chosen degree
The tasks in the jobs the degree leads to, with AI exposure and the share of Claude use on the task that is automation. A task often belongs to several of the occupations; the column shows the main ones. Click a task.
How the list is made
Every occupation has a task list in O*NET, with how important each task is and how many in the occupation do it. A task's weight for a degree is importance × share who do it, weighted by how many with the degree hold the occupation. The ten with the largest weight are shown. AI exposure is a score from 0 to 1 per task, the mean of Eloundou et al.'s three ratings of whether a language model can halve the time the task takes, and the group the score falls in using the same five groups as the occupations: least exposed, low, medium, high, most exposed. Automation is Anthropic's bucketing, directive and feedback loop, as a share of Claude.ai conversations about the task. Tasks Anthropic has no data for are shown without a number.
How Claude is used on the task
The most common jobs for the chosen degree
The ten occupations most people with the degree work in, with exposure and automation.
Method
The page joins four sources in one chain: from degree to occupation, from occupation to tasks, and for each task exposure and Claude use. Institutions have been taken off the page and will come as a separate analysis.
From degree to occupation
The starting point is utdanning.no's register statistics, built on Statistics Norway's registers: for each education code in NUS2000, the number of residents aged 20–70 who hold the code as their highest completed education, split by occupation (STYRK-08) for those employed. Codes are summed to the three-digit group in NUS2000: the first digit is level, the next two are field. Groups with fewer than 100 employed persons with an exposure score are left out. The link is from November 2024. Recent graduates are those who completed one to three years ago, a column utdanning.no does not document.
From occupation to tasks
O*NET is the US Department of Labor's occupational database. Every occupation has a task list, and for each task a sample of incumbents has said how important it is (1–5) and whether they do it. Norwegian STYRK-08 codes go to US SOC codes through the same crosswalk as the exposure measures (ISCO-08, BLS). Where O*NET has specialties under a SOC code, the base code is used. A task's weight for a degree is importance × share who do it, split equally between the SOC codes a STYRK occupation maps to, and weighted by the share of the employed who hold the occupation. O*NET version 30.1.
Exposure per task
Eloundou et al. (2024) had GPT-4 and humans rate every O*NET task: E1 when a language model alone can halve the time the task takes at the same quality, E2 when that needs software built on the model, E0 otherwise. The score in the table is E1 = 1, E2 = 0.5, E0 = 0, the same "beta" weighting as the occupation measure on the index page. The degree's task score is the weighted mean over all tasks in its occupations, not only the ten shown.
Claude use per task
The Anthropic Economic Index classifies a sample of conversations by O*NET task and by five interaction types. Two are automation: the person asks for a finished result (directive), or the model does the work and the person relays what happened (feedback loop). Three are augmentation: task iteration, validation and learning. The numbers here are the latest Claude.ai (chat) sample. Tasks are matched on text. Anthropic publishes only tasks with use, so about a third of the tasks in the tables have a number. The degree's automation share is weighted by each task's share of all Claude.ai use.
Caveats
- Task texts are O*NET's English statements and describe US occupations.
- The register is national per education code. It does not distinguish institutions.
- Claude conversations are global and count conversations, not working time.
- Eloundou's labels are assessments of what is possible, not observed use.
- Small groups get thin numbers. Groups with fewer than 100 employed are left out.
Sources: utdanning.no · O*NET · Eloundou et al. (2024) · Anthropic Economic Index.