How will AI automation hit — like a crashing wave or a rising tide?

The conventional fear about artificial intelligence and jobs is that disruption will arrive suddenly — a wave that knocks workers over before they have time to adapt. But new research from MIT FutureTech, an interdisciplinary research group formed by the MIT Computer Science and Artificial Intelligence Lab and the MIT Initiative on the Digital Economy, suggests a different future.

Drawing on more than 60,000 worker evaluations of AI responses to more than 6,000 text-based workplace tasks, MIT FutureTech researchers asked whether AI capabilities improve in sudden leaps or more gradually.

Their answer points firmly to the latter: Rather than finding crashing waves that suddenly enable AI to take on new tasks, they found a rising tide: AI capability improved broadly across tasks of different lengths rather than surging on a narrow set of tasks.

“You can think of it like standing on a beach,” said Matthias Mertens, a research scientist at MIT FutureTech and lead author of the study. “In a crashing wave scenario, a few people suddenly get knocked over. Instead, we see the water rising around everyone, step by step.”

How the researchers tested AI

Unlike earlier studies that relied on coding benchmarks or other deterministic tests, the research examined how AI performed on representative, real-world workplace tasks.

They identified more than 6,000 text-based tasks from the U.S. Department of Labor’s O*NET database and asked expert workers to evaluate AI-generated outputs for those tasks. In total, the study drew on more than 60,000 evaluations of whether the AI responses were minimally sufficient, average, or better than average without edits.

AI capability high and improving

“Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks” was co-authored by nine researchers from the MIT FutureTech lab, which is directed by principal research scientist Neil Thompson. The team found that AI capability is both high and improving across tasks of very different lengths.

Across all models in the study, AI could already do a minimally sufficient job on roughly 50% to 75% of text-based tasks without requiring edits. And when tasks became 10 times longer, AI success rates fell by only about six to seven percentage points, which suggests that progress is broad rather than concentrated on a narrow set of tasks.

The pace of improvement has also been brisk. Between the second quarter of 2024 and third quarter of 2025, frontier AI models went from achieving a 60% success rate on tasks that take humans 1.5 hours to complete to achieving a success rate of over 70%. 

Overall, the researchers found that the rate at which AI fails to complete a task halves every 2.2 – 2.8 years.

AI Executive Academy

In person at MIT Sloan

Register Now

What it means for workers

The gradual pace of improvement means AI is likely to become capable across many workplace tasks before it reaches near-perfect performance, giving workers time to anticipate potential task-level AI automation.

The researchers estimated that most text-based tasks could reach success rates of 88% to 97% by 2030 at a minimally sufficient quality level, while near-perfect performance remains several years further away.

That gives governments, organizations, and workers a window in which to plan and respond. “The key point is that this gives us time,” Mertens said. “Because the change is gradual, there is a window to adapt.”

Rather than assuming that AI will affect every occupation in the same way, workers and organizations will need to understand how it applies to their own mix of tasks, Mertens said.

That’s important, because the researchers found that AI’s impact varies across different kinds of work.

Success rates ranged from below 50% for text-based legal tasks to more than 70% for text-based tasks associated with installation, maintenance, and repair, suggesting that AI capability is advancing at different rates across occupations. That means the window for adaptation is likely to be wider in some domains than in others.

Some time to prepare

The researchers also caution against reading the findings as a road map for immediate automation. “The success rates should not be interpreted as implying that a corresponding share of tasks can or should be automated today,” Mertens said.

In the study, AI was given all the information needed to complete each task, whereas in practice, that information may be incomplete, difficult to integrate, or subject to regulatory constraints — all of which challenge the feasibility of AI adoption in real-word settings.

In their paper, the researchers also stress that their projections assume that AI will continue improving at its recent pace and should be viewed as an upper-bound scenario. Near-perfect performance is still likely to take considerably longer, particularly where errors carry significant consequences.

Together, those caveats reinforce the study’s central message: Governments, organizations, and workers have time to prepare for automation to avoid being blindsided by it.

“By the end of the decade, these models will be extremely capable on many tasks, but that does not mean entire occupations will simply disappear,” Mertens said. “Many jobs include a subset of tasks beyond the reach of language models, so work is more likely to be reorganized between humans and AI.”

“Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation From Thousands of Worker Evaluations of Labor Market Tasks” was written by Matthias Mertens, Adam Kuzee, Brittany S. Harris, Harry Lyu, Wensu Li, Jonathan Rosenfeld, Meiri Anto, Martin Fleming, and Neil Thompson.

Matthias Mertens is a research scientist at MIT FutureTech. His primary interests include firm productivity, market power, wage determinants, and reallocation processes. Currently, he is studying the impacts of information technology advancements on firm productivity and the determinants of increasing AI capabilities.