Will AI change your job? What OpenAI's labor study found

Last Friday, four researchers posted a paper on arXiv with a playful title: GPTs are GPTs. The first "GPTs" means generative pre-trained transformers, the family of models behind ChatGPT. The second means general-purpose technologies, the economists' term for inventions like electricity or the steam engine that end up touching almost every kind of work. Three of the authors are affiliated with OpenAI (one also with OpenResearch); the fourth, Daniel Rock, is at the University of Pennsylvania.

Their question is simple to state: if workers had access to a capable language model, how much of their day would it affect? Their headline answer has been widely repeated since Monday. Roughly 80% of the U.S. workforce could see at least 10% of their tasks affected, and around 19% could see at least half of their tasks affected.

How the researchers measured it

The study does not look at jobs as single blocks. It breaks them into tasks using O\*NET, a U.S. government database that describes 1,016 occupations through about 19,000 specific tasks. A data analyst, for example, cleans data, writes reports and meets clients; each of those is rated separately.

For each task, raters asked one question: would access to a GPT model, or software built on one, cut the time needed to do this task by at least 50% while keeping the quality the same? The paper sorts tasks into three groups:

Both humans and GPT-4 itself did the rating, and the authors report similar levels of agreement between them. They then rolled the task ratings up to each occupation.

Who is most exposed

The pattern surprises many readers. Higher-wage jobs tend to show more exposure, not less, which the authors note runs against earlier studies of machine learning's reach. Jobs that lean on programming and writing skills score high; jobs that depend on science and critical thinking score lower. Vice's summary of the paper names writers, journalists, accountants, tax preparers, mathematicians, web designers and legal secretaries among the most exposed. By industry, information processing ranks high, while manufacturing, farming and mining rank low.

A worked example: rating your own job

You can apply the paper's question to your own work in about fifteen minutes. List your tasks, then ask an assistant to help you think through each one, while you stay the judge.

I work as a [your job]. Here are eight tasks I do in a typical week:
1. ...
2. ...
For each task, tell me:
- whether a chatbot alone could help me finish it in half the time at the same quality,
- whether it would need extra software (for example, access to my files or calendar),
- or whether it would not save real time.
Explain your reasoning in one sentence per task. Mark anything you are unsure about.

Treat the answers as a starting point. You know your job better than any model; the value is in seeing your work as a list of tasks, some of which may change sooner than others.

How should you read the 80% and 19% figures?

Start with the word "exposed." The authors define exposure as a stand-in for potential economic impact, and they say plainly that it does not separate tools that help workers from tools that replace them. A task that takes half the time could free you for other work, raise your output, or reduce the number of people needed. The paper does not predict which.

Next, notice the thresholds. The 80% figure only requires that one task in ten be exposed. That bar is low; it covers someone who uses a chatbot to draft the occasional email. The 19% figure is stricter, needing half of a worker's tasks to be exposed, and it counts both the model alone and tools that others might build on top of it. When human raters counted only what the model can do by itself, with no extra software, the share of workers with more than half their tasks exposed was about 3%. Counting future tools and other generative models, it rose as high as 49%. The spread tells you how much depends on software that does not exist yet.

Then consider who wrote it. Vice points out that OpenAI, which sells these models, produced the study. The paper itself lists limitations: the human raters knew the models well and did not work in most of the jobs they scored, and GPT-4's ratings change with the wording of the prompt. The authors also stress that being technically possible does not guarantee that companies will adopt a tool, or that productivity will rise.

So the fair reading is this: most jobs contain some tasks that language models can speed up, and a meaningful minority contain many. Whether that brings better jobs, fewer jobs or both depends on choices by employers, workers and policymakers that the study does not measure. Learning which of your own tasks fall into which group is a practical first step.