Assessing the Real Impact of Automation on Jobs

Who’s more at risk of being replaced by automation: a crossing guard or an air traffic controller? Both prevent collisions, but only one requires years of specialized training. An air traffic controller could do a crossing guard’s job – but not the other way around. And while the guard’s work is untouched by automation, the controller’s tools are increasingly automated to reduce human error.

The answer, says MIT economist David Autor, is nuanced. In a talk hosted by the Stanford Digital Economy Lab at the Stanford Institute for Human-Centered AI, Autor challenged the assumption that automation exposure simply means job loss. “Exposure is not a very useful term,” Autor said. “Is it the case that if you’re exposed, you’re hosed?” 

Not necessarily, he said. He pointed to Uber: Wages for taxi drivers stagnated, but employment rose 249% from 2000 to 2020 as automation lowered the barrier to entry. In contrast, proofreaders saw wages rise but job numbers decline as automation removed simpler tasks while adding expert tasks that made the role more specialized. 

“Proofreading used to mean spell-checking. Now it’s about helping people write,” Autor said.

In short, Autor found that automation both replaces and augments expertise – it depends on whether rote tasks are removed and expert ones added, and how specialized a role becomes as a result.

An Objective Model to Track Automation’s Impact

To evaluate the impact of automation on occupations, Autor tracked the addition and removal of tasks in job descriptions from 1977 to 2018, alongside shifts in wages and employment. To determine what constituted an expert versus an inexpert job, Autor drew from Zipf’s law and the efficient coding hypothesis. These concepts explain how language evolves to include common shortcut words in order to streamline communication. Initially, these jargon words are understood by a select few experts, but they enter the common vernacular over time. 

“I will be able to say ‘LLM’ or ‘GPT’ or whatever, and you’ll all know what I mean,” Autor said. “You wouldn’t have known what I meant five years ago, but you know what I mean now.”

Autor used this idea to distinguish between common – or nonexpert – words and less frequent – or expert – words as he evaluated job descriptions between 1977 and 2018.

Impact of Exposure Depends on Expert Supply

Autor found that for jobs that gained inexpert tasks but lost expertise, wages declined as technology made it possible for more people to do more of the tasks required – like in the Uber scenario. Knowing the most highly trafficked places to pick up passengers and the right routes to get them to their destinations were no longer expert tasks – more people could do the job. Taxi driving became less specialized.

Autor’s model found that the opposite trends occurred – employment went down and wages rose – for jobs that lost inexpert tasks but gained expert tasks that upgraded their expertise levels. In the proofreader example, inexpert tasks like “places proof and copy side by side on reading board” went away, but “consult reference books to check references with rules of grammar and composition” were added.

“Expertise is much closer to a supply change,” he added. “When expertise falls, it’s a reduction in barriers. When expertise requirements rise, it’s an increase in barriers.”

Predicting a Job’s Future by Assessing Routine Tasks

Autor used a large language model to classify tasks into three categories: abstract tasks that require creativity, reasoning, and interpersonal skills; routine tasks that follow clear, repetitive rules; and manual tasks that involve physical effort and common sense but little formal training. He found that 64.5% of removed tasks in his data set were routine, while 75.6% of added tasks were abstract. In other words, jobs with many routine tasks in 1977 had far fewer routine tasks by 2018.

What does that imply for a job’s expertise level? In some occupations, losing routine tasks led to lower wages. In others, it increased specialization and pay.

“For some things you’re taking away the supporting activities. You’re allowing people to specialize and focus on their comparative advantage,” Autor said. “For other sets of occupations, you’re taking away their primary activity, and so you’re removing what’s special about that occupation and reducing it down to the generic activities that many more people could do,” he said.

Autor noted that these results show the “exposure paradox” in action. “These are exposed occupations, but the exposure has completely different meanings for how that work is going to change,” he said.

Does Automation Replace Experts or Augment Expertise? The Answer Is “Yes.”

Returning to his study’s central question, Autor said automation both replaces experts and augments expertise. It depends on whether abstract or routine tasks are removed and added and whether expertise continues to be in demand for a given role.

He emphasized that focusing only on exposure misses the point. What matters is the one-way fungibility of expertise: Experts can do nonexpert work when displaced, but not the reverse. The key question is whether expertise can transfer – or whether it becomes stranded in an increasingly routine domain.

AI’s Impact Is an Open Question

Autor acknowledged that his framework can’t fully predict AI’s effects, since his data ends in 2018 – before tools like ChatGPT launched broadly to the public. And because it’s still unclear which tasks AI will eliminate or enhance, the long-term impact on expertise is unknown.

Still, he noted that AI could help nonexperts access skills they previously lacked. Paralegals and nurse practitioners, for example, could gain from lower barriers to expertise, he said.

“In the good scenario, AI enables more people to do more expert tasks,” Autor said. “Now that’s not good for experts, but it’s good for a lot of other people.”

Watch the full Digital Economy Lab seminar with David Autor.