AI shows huge gains in labs, but a study of 25,000 real workers found it has almost zero effect on their actual output.
March 19, 2026
Original Paper
Generative AI and Worker Productivity: A Systematic Review and Quantitative Evidence Synthesis (2023-2026)
SSRN · 6366218
The Takeaway
This 'empirical paradox' exists because lab studies measure isolated tasks where AI excels, but in the complex, multi-tasking environment of a real job, those gains evaporate. It suggests that the widely reported '40% productivity boost' from AI is a laboratory artifact that hasn't yet manifested in the actual economy.
From the abstract
The only study using nationally representative administrative records-tracking 25,000 workers across two years following the public release of ChatGPT-finds a confidence interval ruling out earnings effects larger than two percent, yet controlled experiments conducted over the same period report productivity improvements of 14 to 55 percent. Resolving this empirical paradox is the central aim of this systematic review and quantitative evidence synthesis of studies examining generative artificial