Guide · Research
What the research really says about AI and jobs
A plain-English tour of the main studies on AI and work, what they found, what they didn't, and how to read headlines about AI taking jobs.
30 October 2026 · 3 min read
Headlines about AI and jobs swing between “AI will take your job” and “nothing to worry about”. The research is more careful than either. Here is what some of the most-cited work actually found, in plain English, and how it informs the AI Survival Score.
Exposure is not the same as replacement
Most studies measure exposure: how much of a job’s tasks AI could, in principle, do or speed up. That is not the same as predicting job losses. A task can be exposed and still be done by people for reasons of cost, trust, regulation, quality or simple preference. Exposure can also mean augmentation: the person stays, and gets faster.
Keep that distinction in mind whenever you read a big number.
“GPTs are GPTs” (OpenAI and the University of Pennsylvania, 2023)
This widely cited paper by Eloundou, Manning, Mishkin and Rock rated the tasks in hundreds of occupations by whether a large language model could cut the time to do them significantly. Its headline estimate: around 80% of the US workforce could have at least 10% of their tasks affected, and around 19% could see at least half of their tasks affected.
Two findings stood out. First, exposure was higher for many higher-wage, white-collar jobs, the reverse of earlier automation waves that mostly hit routine manual work. Second, tasks involving science and critical thinking skills tended to be less exposed than tasks centred on writing and programming.
The International Labour Organization (2023, with later updates)
The ILO’s global analysis of generative AI and jobs looked at exposure across occupations worldwide. Its main conclusion was that the most likely effect for most jobs is augmentation rather than automation: AI takes on some tasks while people keep the job and do more of the rest.
The occupation group with the highest exposure was clerical support work, such as data entry, typing and routine administration. Effects also differ by country, because the mix of jobs differs: economies with more office work are more exposed.
The Anthropic Economic Index (2025 onwards)
Rather than asking what AI could do, this research looks at what people actually use an AI assistant for, based on privacy-preserving analysis of real conversations. Early reports found usage concentrated in software development and writing tasks, and that more conversations looked like people collaborating with AI (augmentation) than handing work over entirely (automation).
Explore the Anthropic Economic Index.
What the studies agree on
- Tasks, not jobs. Almost every serious study works at the task level, because jobs are bundles of tasks with very different exposure.
- Writing, coding, data handling and routine admin are among the most exposed tasks.
- Physical work, care, and work built on trust and accountability are among the least exposed.
- Augmentation is common. Many people will keep their jobs and see them change.
What they can’t tell you
No study can tell you what will happen to your job. They work with averages across whole occupations, they estimate what is technically possible rather than what employers will do, and AI capabilities keep moving.
They also can’t measure the part that is most in your control: how quickly you adapt. That is why the AI Survival Score combines task exposure with your AI adoption, learning habits and resilience.
How we use this research
Our scoring model was informed by research like the above, but its calibration is our own, not numbers copied from any study, and none of these researchers endorse it. The methodology page explains how it works and its limitations.
Want to see where your own tasks fall? Take the free test, or browse how AI affects 16 common tasks.
Keep reading
What's your AI Survival Score?
12 questions, about 2 minutes, no sign-up.
Take the test