Artificial intelligence (AI) may reshape many industries, but the impact of the nascent technology on various jobs remains unclear. Daniele Quercia and colleagues used machine learning to investigate itself, by identifying patents for AI technologies that may impact various occupational tasks. The work is published in PNAS Nexus.

The model used a dataset of 17,879 task descriptions from O*NET, a US government-run occupations database, as well as 24,758 AI patents filed with the United States Patent and Trademark Office between 2015 and 2022 and measured semantic similarity between occupation task descriptions and patent descriptions. The analysis was not merely an exercise in word matching but compared entire task descriptions to entire patents.

For each task, the most similar patent was identified and if the patent was more than 90% similar to the task, that task was considered impacted by AI. Each occupation was then given an AI Impact (AII) score by dividing the number of its tasks impacted by AI patents by the total number of tasks for that occupation.

Using this method, the team identified the most impacted occupations, which include orthodontists, security guards, and air traffic controllers. The team also identified the least impacted occupations, which include pile driver operators, dredge operators, and graders of agricultural products.

According to the authors, occupations including repetitive tasks were not always those most impacted by AI—jobs that include tasks in a specific sequence that produce a machine-readable output were most likely to be impacted.

The team found that research often overestimates how much AI will take away jobs because sectors potentially impacted by AI, like health care and transportation, currently need more workers, not fewer and because AI is likely to augment rather than replace many jobs. For example, AI can analyze medical images like X-rays to help doctors find issues but cannot replace doctors.

According to the authors, tech developers should focus on creating tools that support rather than replace people, and leaders should promote the use of AI with an emphasis on training and education so everyone can benefit.

Provided by PNAS Nexus