Productivity
Productivity is the amount of output produced per unit of input. Labor productivity divides output, such as value added, by hours worked; total factor productivity (TFP) is the residual growth in output not explained by growth in labor and capital, often read as a measure of technology and organizational efficiency. Because long-run living standards depend heavily on productivity growth, any technology claimed to change how much people can produce, AI included, is judged against it. The concept is simple but measurement is not: an hour of customer service, a line of software or a drafted memo has no natural unit of output.
The Solow paradox is the field's founding puzzle about technology and measured output. In a 1987 New York Times book review, economist Robert Solow wrote that "you can see the computer age everywhere but in the productivity statistics." Computing power and investment were soaring, yet US measured productivity growth was sluggish. Erik Brynjolfsson sharpened the question in his 1993 article "The Productivity Paradox of Information Technology." Explanations include mismeasurement of quality and variety, benefits concentrated in a few firms, and long lags. Paul David's 1990 paper "The Dynamo and the Computer" noted that factories took decades after the electric dynamo arrived to reorganize around electric motors, which is the historical template for lags. US productivity growth did pick up in the late 1990s, and many analysts connected it to information technology.
The productivity J-curve explains why general-purpose technologies can first look unproductive. Brynjolfsson, Rock and Syverson formalized it in "The Productivity J-Curve" (2021): adopting a technology like AI requires large intangible investments — reorganized processes, new skills, data pipelines, complementary software — that are not counted as output, so measured productivity may flatten or even dip at first, then rise as those investments pay off. The pattern also implies that early measured gains understate eventual ones. It likewise implies the reverse caution: a strong early result in a lab or pilot does not guarantee a similar firm-wide or economy-wide gain, since integration costs, error handling and demand constraints sit between the two.
Controlled experiments on AI tools provide the best direct measurements, and they show real but task-specific gains. In Peng, Kalliamvakou, Cihon and Demirer's 2023 study of GitHub Copilot, developers asked to implement an HTTP server in JavaScript finished the task about 55.8 percent faster with Copilot than the control group. Noy and Zhang's 2023 Science experiment on professional writing tasks found AI assistance reduced time and raised rated quality, and Brynjolfsson, Li and Raymond (2023) found roughly 14 percent more issues resolved per hour for support agents, with the largest gains for novices. Dell'Acqua and colleagues' 2023 field experiment with consultants described a jagged frontier: on tasks inside what the model did well, performance improved considerably, but on a task deliberately outside it, consultants using AI were less likely to reach the correct answer. A 2025 randomized trial by METR on experienced open-source developers working in familiar codebases found that they took longer with AI tools, even though they expected to be faster.
Caveats apply to every one of these numbers. The tasks are bounded and often chosen because they are well suited to the tool; participants are volunteers or employees at one firm; and speed is not value, since faster code that needs more debugging, or text that is fluent but wrong, can offset the time saved. Gains at the task level also need not translate to output at the firm level if the time saved is absorbed by meetings, if the bottleneck is elsewhere, or if the product's demand does not grow. Macroeconomic estimates are correspondingly cautious and widely spread, depending on assumptions about the share of tasks affected and the cost savings on each.
The usual pitfalls are quoting a percentage without its denominator and treating productivity as a synonym for job cuts. "55.8 percent faster" refers to one task in one experiment, not to developers' overall output. Higher productivity can mean more output with the same workers, the same output with fewer, or higher quality, and which one occurs depends on demand, competition and management choices. Productivity gains also say nothing alone about who receives them; wages, profits and prices share the benefit in proportions that vary by economy and era.
Read productivity as a ratio that is easy to state and difficult to measure, where promising task experiments show what a tool can do, the J-curve explains why the statistics lag, and the Solow paradox is the reminder to check actual output before declaring either triumph or failure.