AI didn't kill jobs from 2010 to 2023. It changed what workers do.

Artificial intelligence is often portrayed as a job killer. A new study from MIT Sloan tracking AI adoption from 2010 to 2023 tells a different story. AI’s impact is task-level, not occupation-level. And firms that adopt AI grow faster, hiring more workers overall.

The study, co-authored by MIT Sloan associate professor Lawrence Schmidt, Menaka Hampole of Yale, and Dimitris Papanikolaou and Bryan Seegmiller of Northwestern, used natural language processing to construct measures of workers’ task-level exposure to AI and machine learning. The dataset captures variation across firms, occupations, and time. The results challenge both the alarmist and the utopian narratives.

Here is what the data actually shows.

AI-exposed roles did not shrink overall. From 2014 to 2023, workers in roles with high AI exposure saw their share of total employment grow by about 3% over five years. That is counterintuitive. If AI automates tasks, why would employment in those roles increase? The answer is firm-level productivity. Companies that adopted AI grew faster. That growth created enough new demand to offset any task-level displacement.

The mechanism matters. When AI can perform most of the tasks in a job, the share of people in that role within a firm falls by about 14%. But when AI’s impact is concentrated in a few tasks, leaving other responsibilities untouched, employment in that role can grow. Workers shift their time to activities where AI is less capable — critical thinking, novel problem-solving, interpersonal judgment.

Firms that use AI extensively are larger, more productive, and pay higher wages. A large increase in AI use correlates with about 6% higher employment growth and 9.5% more sales growth over five years. These are not small effects. They suggest that AI is not a zero-sum technology. It is a productivity technology that expands the pie.

The researchers found that exposure to AI is greatest in high-paying roles involving information processing and analysis. That is a reversal from earlier technology waves. Computerization and factory automation in the 1980s through early 2000s mainly displaced middle-skill routine jobs — clerical work, basic bookkeeping. AI targets the top of the wage distribution.

Which jobs shrank and which grew tells a revealing story. Top-paying roles like management analysts, aerospace engineers, and computer and information research scientists saw employment within firms fall by about 3.5% over five years. But those roles are concentrated in firms that adopt AI. The productivity gains those firms realize more than offset the losses. The highest-paid jobs still make up a slightly bigger share of overall employment than less exposed roles.

Business, financial, architecture, and engineering jobs shrank by about 2% to 2.5% over five years. These occupations have a high share of tasks that match what AI can do. Business and financial jobs are more exposed than architecture and engineering, but they are also more likely to be in heavy AI-adopting firms, which helps offset some losses.

Legal jobs gained the most. They face little direct impact from automation and are often in firms that use AI heavily, leading to a predicted 6.4% increase in employment.

Even low-exposure jobs are not safe. Food service jobs wilted relative to other roles not because AI can do the work but because employers that do not use AI grow more slowly, reducing demand for workers. The same dynamic applies to any occupation in a non-adopting firm. Workers in roles amenable to AI but employed by non-adopting firms are not better off. Their employers grow less rapidly than AI-adopting peers.

The study’s dataset runs to 2023, so most of it predates the rapid rise of generative AI tools like ChatGPT, which launched in late 2022. Schmidt calls this the million-dollar question. Generative AI can learn from fewer examples and take on a wider range of tasks, potentially leaving fewer for humans to shift to. It could reduce the gains from task reallocation. It could also make everyone more productive — a rising tide that lifts all boats.

The research makes one thing clear: management decisions shape AI’s impact more than the technology itself. Schmidt emphasizes task reallocation. Take the existing workforce, work with the AI, and reallocate time toward tasks where people have a comparative advantage. Do not wait for a top-down rollout. Give employees hands-on access to tools now. Pick the right versions and features for the work. Use AI not just to save time on routine work but to tackle complex problems and generate fresh ideas.

The data from 2010 to 2023 does not support the job-killer narrative. It supports a task-transformation narrative. AI changes what workers do, not whether they work. The firms that figure out task reallocation will grow. The firms that do not will shrink. That is the bet.