Large language models could change parts of work across a wide range of occupations, according to a study by OpenAI, OpenResearch, and the University of Pennsylvania. The researchers estimate that at least 10 percent of the work of about 80 percent of U.S. workers could be affected. For 19 percent of workers, the models could influence at least 50 percent of tasks.
What the study measures
The team examined 19,265 tasks and 2,087 work processes in the O*NET 27.2 database. It covers 1,016 occupations and describes their tasks and work processes, including detailed work activities such as “study scripts to determine project requirements.”
To estimate exposure to AI, the researchers focused on how much time a language model or GPT-based system might save when completing a task. A task counted as directly exposed if a GPT model or a system built on GPT could cut the time needed by 50 percent.
That threshold is a way to make potential effects measurable. The researchers describe the choice as “somewhat arbitrary,” while arguing that substantial productivity gains could make adoption more likely and help human raters interpret tasks consistently.
Exposure does not mean replacement
The study’s measure is about a task’s potential exposure to language models. It does not distinguish between AI helping a person do the work and AI taking over the work. A job can therefore be exposed under the study’s definition without the findings showing that a worker will lose that job.
To assess exposure in the U.S. labor force, the researchers used human raters and GPT-4. Their assessment considered both the capabilities of current models and those of anticipated GPT-based software. The result is an estimate of where these systems might affect work, rather than a forecast of exactly how employers or workers will use them.
Writing and programming stand out
The study finds that higher-paying jobs are more affected overall. Programming and writing jobs are particularly exposed, while jobs that depend heavily on science and critical thinking are less at risk, according to the researchers.
Occupations with high barriers to entry also appear more likely to be affected than those with low barriers. At the other end, occupations with no exposure in the study are mainly manual jobs, including stonemason and cook. These roles can often be learned with less schooling up front than occupations requiring many years of study.
That difference does not mean manual work is insulated from all consequences. The researchers suggest an indirect possibility: if language models make many office jobs obsolete, craft occupations could attract more workers, increasing competition and performance pressure. This is presented as a possible knock-on effect, not a measured outcome.
Why the effects could keep spreading
The researchers connect GPTs, meaning Generative Pre-trained Transformers, to the idea of general-purpose technologies. The same acronym also refers to technologies such as electricity, fire, or mobility that can have broad economic effects. Their comparison is a hypothesis that language models may share some characteristics with those technologies.
The study also argues that economic effects could continue to grow even if development of new model capabilities stopped today. Complementary technologies could expand the range of potential impact. This points to a wider question than what a model can do on its own: how it might change work when integrated into other tools and processes.
The researchers say further work is needed to understand whether language models augment or replace human labor, and what they mean for work quality, inequality, and skills development. Those questions matter for workers and for policymakers considering the role of AI in the future of work.