AI tools are being used widely in some professions but have not yet caused widespread job losses, according to new research analysing how artificial intelligence is being adopted in the workplace.
The study introduces a metric called "observed exposure", which measures how frequently tasks in different occupations are actually being performed using large language models rather than simply assessing whether the technology could theoretically automate them. The approach combines data on AI usage with occupational task data to track where automation is occurring in practice.
Early findings suggest that AI adoption remains well below its potential scope and has not yet produced a measurable rise in unemployment among workers in the most exposed occupations.
Measuring exposure
The research combines three main sources of information to estimate AI exposure across jobs. These include occupational task descriptions from the O*NET database, task-level estimates of what large language models can theoretically perform, and real-world usage data from AI systems.
The resulting measure focuses on tasks that are both technically feasible for AI and already being carried out using such tools in work-related contexts. Automated uses receive greater weight than cases where workers simply use AI to assist with tasks.
The analysis indicates that the gap between what AI could theoretically perform and what it is currently used for remains large. For example, tasks in computer and mathematics occupations are considered highly feasible for automation, yet current AI usage covers only about one third of those tasks.
Some tasks that are theoretically automatable may still face practical barriers to adoption. These include legal restrictions, workflow requirements, the need for verification by humans, or the integration of specialised software systems.
Most exposed jobs
The study identifies several occupations with relatively high levels of observed exposure to AI.
Computer programmers rank highest, with about 75% of their tasks considered exposed under the measure. Customer service representatives follow with around 70%, while data entry keyers show roughly 67% exposure due to the automation of document processing and data input tasks.
Other highly exposed roles include medical record specialists, market research analysts, financial analysts, and software quality assurance testers.
By contrast, about 30% of workers fall into occupations with no measurable exposure in the dataset. These include jobs such as cooks, motorcycle mechanics, lifeguards, bartenders and dishwashers, where tasks rarely appear in the AI usage data.
Many of these roles involve physical activity, specialised equipment or interpersonal interaction that remains difficult to automate.
Growth outlook
The analysis also compares AI exposure with employment projections for the coming decade.
Using forecasts from the US Bureau of Labour Statistics covering the period from 2024 to 2034, the researchers find that occupations with higher exposure tend to have slightly weaker projected growth.
For every ten percentage point increase in observed AI exposure, projected job growth declines by about 0.6 percentage points. The relationship is modest but suggests that occupations where AI is already used more extensively may see slower expansion in the future.
Workers in the most exposed occupations also show distinct demographic patterns. They are more likely to be older, more educated and higher paid than workers in occupations with little or no AI exposure.
The data indicate that the highly exposed group earns about 47% more on average and includes a much higher share of workers with graduate degrees.
Employment trends
Despite these differences in exposure, the research finds no clear evidence that AI has yet caused higher unemployment among workers in the most exposed occupations.
Comparisons of unemployment rates between highly exposed workers and those in low-exposure roles show little divergence since the widespread introduction of generative AI tools. Trends for both groups have moved broadly in parallel in recent years.
The analysis focuses on unemployment as a primary indicator because it reflects workers who want jobs but cannot find them, making it a direct signal of potential economic harm.
The results suggest that, so far, AI adoption has not significantly displaced workers on a large scale.
Younger workers
One area where early changes may be emerging is hiring among younger workers.
The study finds tentative evidence that job entry rates for workers aged 22 to 25 have declined in occupations with higher AI exposure. Data tracking monthly job starts show that the rate at which young workers move into these occupations has fallen by roughly half a percentage point since the release of generative AI tools.
On average, this represents a decline of about 14% in the job-finding rate for these roles compared with levels recorded in 2022.
Researchers note that the finding is only marginally statistically significant and may reflect several factors beyond automation. Young workers who are not entering these roles may remain in existing jobs, shift to other occupations or pursue further education.
The study concludes that current labour market data do not yet show a broad disruption from AI adoption. However, tracking how task coverage evolves as AI systems become more capable may help identify future shifts in employment patterns.