Earlier computers automated tasks someone could write as rules. Modern AI systems learn patterns from huge amounts of data, so they can do some things no one ever wrote rules for: drafting a letter, summarizing a contract, writing code, translating a menu. This reaches into exactly the nonroutine cognitive work that Level 3 described as hard to automate. Here the question is what AI means for jobs.

A different pattern of exposure

Researchers say a task is "exposed" to AI when AI could do it, or cut the time needed to do it, by a large amount. Exposure means that AI could do or greatly speed up a task, not that the job will disappear. An exposed task can be automated, with AI doing it instead of a person, or augmented, with AI helping a person do it better or faster. The same tool can play either role depending on how employers and workers use it.

The striking finding is who is exposed. Robots and older software hit factory and clerical work hardest. Generative AI exposure is highest in well-paid occupations that require more education, such as programming, writing, finance, law and customer support, while plumbers, cooks and care workers are among the least exposed.

Competing forecasts

Estimates of the damage have varied wildly, mostly because of method. In 2013 Carl Benedikt Frey and Michael Osborne at Oxford had experts label a sample of whole occupations as automatable or not, then used a statistical model to rate 702 occupations. Frey and Osborne estimated that about 47 percent of US jobs were at high risk of automation. The number was quoted everywhere.

In 2016 Melanie Arntz, Terry Gregory and Ulrich Zierahn, writing for the OECD, repeated the exercise using surveys of what individual workers actually do. They found only about 9 percent of jobs at high risk in the countries they studied. Task-based estimates came out much lower because workers in the same occupation do different mixes of tasks, and many do tasks that are hard to automate even in jobs that look exposed.

Newer studies focus on generative AI. A 2023 paper by Tyna Eloundou and coauthors from OpenAI and the University of Pennsylvania estimated that about 80 percent of US workers could have at least 10 percent of their tasks affected by large language models, and about 19 percent could see at least half their tasks affected. In 2024 the International Monetary Fund estimated that almost 40 percent of jobs worldwide are exposed, rising to about 60 percent in advanced economies. The IMF found higher exposure in advanced economies because they have more cognitive, office-based jobs. Poorer countries have less exposure but also less ability to gain from AI.

Optimists and skeptics

Optimists point to evidence that AI can raise productivity and narrow gaps between workers. In a study of about 5,000 customer support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised the number of issues resolved per hour by about 14 percent on average. In the customer support study, the least experienced agents gained the most from the AI assistant, about 34 percent, while top performers gained little. The tool seemed to spread the know-how of the best workers to newcomers.

Skeptics doubt the big numbers. Daron Acemoglu argued in 2024 that only a modest share of tasks can be profitably automated soon, and Acemoglu estimated that AI would raise total productivity by well under 1 percent over a decade, about 0.7 percent. Others reply that he leaves out new tasks and future improvements. Early labor market data are also mixed. A 2025 study of Danish workers by Anders Humlum and Emilie Vestergaard found almost no effect of chatbot use on earnings or hours so far, while a 2025 study of US payroll data by Brynjolfsson and coauthors found falling employment for workers aged 22 to 25 in the most exposed jobs. These findings are very recent and may change as more data arrive.

Worked example: a translation agency

An agency employs twelve translators. Machine translation now produces usable first drafts in seconds. Apply this level. Exposure is high: drafting is most of the job. But which way will the agency use the tool? If it automates, it may keep a few translators as editors and cut the rest, and pay per word may fall. If it augments, each translator may handle more work, take on clients who could not afford human translation before, and spend more time on legal documents and literary texts where errors are costly. Both have happened in the real translation industry, and surveys of freelance translators report falling rates for routine work. The forecasts in this level would each predict something different for the agency. The honest answer is that exposure tells you where change will come, not whether it will mean fewer jobs, lower pay or better work. That depends on choices made by firms, workers and governments, which the rest of the course takes up.