Large-Scale Worker Study Finds AI Automation Rising Broadly Across Jobs, Not in Sudden Capability Spikes
Summary
- • 17,000+ worker evaluations across 3,000+ job tasks show broad, steady AI capability gains.
- • AI completed 3–4 hour tasks at 50% success in 2024-Q2, rising to 65% by 2025-Q3.
- • LLMs projected to handle 80–95% of text-based tasks successfully by 2029.
- • Broad 'rising tide' improvement contradicts METR's 'crashing wave' hypothesis of narrow capability spikes.
Details
O*NET task taxonomy used for broad cross-sector coverage
The study draws on 3,000+ text-based, LLM-addressable tasks from the U.S. Department of Labor's O*NET occupational database, with evaluations conducted by workers actually employed in those roles — lending practical, domain-grounded validity beyond typical benchmark design.
AI success rate on 3–4 hour tasks: 50% (2024-Q2) → 65% (2025-Q3)
This ~15 percentage point gain over roughly five quarters forms the empirical basis for the researchers' 2029 projection. The tasks evaluated require approximately 3–4 hours of human effort, representing substantial, real-world complexity.
80–95% projected success on most text tasks by 2029
Extrapolating current trends, researchers estimate LLMs will achieve 80–95% average success rates on text-related tasks at a minimally sufficient quality level by 2029. Near-perfect performance or meaningfully higher quality output would require additional years beyond that window.
Findings contradict METR's 'crashing wave' model of AI automation
METR's recent work characterised AI capability growth as abrupt and concentrated in narrow task domains. This study finds the opposite: a 'rising tide' of gradual, broad-based improvement lifting many task types simultaneously — with little empirical evidence for wave-like surges.
Economic and labor market disruption will lag AI capability gains
The authors explicitly distinguish what AI can do from what organisations deploy. Even with rapid capability progress toward 2029, actual labor market impact depends on adoption rates, workflow integration, regulatory responses, and workforce transitions — a timeline that could be substantially longer. The paper is an arXiv preprint and has not yet undergone peer review.
Research = study methodology/findings; Stat = specific data point; Insight = analytical conclusion; Context = framing and caveats
What This Means
For business leaders and policymakers, AI-driven disruption to white-collar, text-based work is already underway and broadening — not a future risk. Rather than bracing for isolated capability shocks, organisations should plan for continuously widening AI competence across most knowledge-work functions through the late 2020s. The critical uncertainty is the gap between AI capability and actual deployment, making workforce transition planning and adoption strategy as important as tracking the underlying technology itself.
Sentiment
Balanced reassurance: gradual task-level AI adoption broadening across jobs, not sudden shocks, but calls for urgent workforce planning
“MIT spent a year studying AI and job displacement. The conclusion: not a crashing wave but a rising tide. Half of U.S. jobs reshaped over 2-3 years at the task level, not role level. The executives treating this as a distant abstraction are the ones who will be caught off guard.”
“A study from MIT Computer Science and Artificial Intelligence Laboratory challenges the idea of an imminent AI-driven job collapse, suggesting instead that AI will gradually reshape work rather than replace it outright... Work is being restructured around systems that can assist, accelerate, and sometimes act. AI adoption will move at the speed of trust, not capability.”
“The underlying study, 443 million hours of behavioral data across 163,638 workers, doesn't show AI making work harder. It shows organizations filling freed capacity with more volume instead of redesigning how work is deployed. Email time up 104% post-AI adoption... That's not an AI problem. That's a management operating model that was never designed to govern amplified work.”
“AI JOB APOCALYPSE FALLS APART... Jobs with high AI exposure are not shrinking. They are growing faster and paying better... AI is not replacing workers. It is amplifying them.”
Split
~70/30 positive augmentation and growth vs concerns over management failures and task unbundling