Research by academics at King’s College London and the AI Objectives Institute has shed light on why what matters is not just how much of a job AI can do, but which parts. Dr. Bouke Klein Teeselink and Daniel Carey analyzed hundreds of millions of job postings across 39 countries before and after the release of ChatGPT in November 2022. They found that occupations with a large number of tasks exposed to AI automation, for example basic administration or data entry, saw a 6.1% decline in job postings on average. Importantly, however, this effect depends not only on how many tasks are exposed, but also on which tasks.

When AI automates the routine, less-skilled parts of a job, the work that remains tends to be more specialized. Fewer people can do it, so wages rise. The researchers cite the example of a human resources specialist whose administrative paperwork is now handled by AI, leaving them to focus on complex employee relations and judgment calls.

But when AI can perform the more specialized, cognitively demanding tasks, wages decrease because the job no longer requires scarce expertise. This example can apply to roles such as junior software engineers, the researchers found.