Does generative AI productivity actually improve or decline in practice?

Auste Simkute, Lev Tankelevitch, and colleagues at the University of Edinburgh, Microsoft Research, and University College London argue that generative AI's productivity gains are not guaranteed, writing that "while they boost productivity in some studies, many others show that users are working ineffectively with GenAI systems and losing productivity." Rather than treating this as a new phenomenon, they trace it to three decades of Human Factors research on automation in aviation, automated driving, and industrial plants, and name "four key reasons for productivity loss with GenAI systems: a shift in users' roles from production to evaluation, unhelpful restructuring of workflows, interruptions, and a tendency for automation to make easy tasks easier and hard tasks harder."

The mechanism they invoke is the "ironies of automation" — the long-standing Human Factors observation that "the more advanced an automated system is, the more important the human operator may be," because automation displaces manual control but still leaves a human to supervise it, often without adequate support. Applied to GenAI, the authors say users' "roles have shifted from producing output to evaluating it, often with little contextual information and situational awareness," a burden worsened because GenAI "produce[s] outputs at a capacity too demanding for adequate evaluation, with questionable reliability, and with poor explainability." Separately, "poor system and interface design can result in unhelpful restructuring of workflows," interruptions from AI suggestions break users' flow (echoed in programmers "losing their state of flow when interrupted by long code suggestions"), and automation's selection of which tasks to take over produces what they call "task-complexity polarization," also known as "clumsy automation" in Human Factors research — easy tasks get easier, hard tasks get harder.

This gives the evaluation-burden side of a cost that Can generative AI reduce overhead work that adds no value? names from the input side: where Larsen-Ledet's "mundane articulation" is the unrewarded context-setting work users do before generation, Simkute and colleagues describe the unrewarded scrutiny work users do after it — together the two papers bracket a GenAI task with cost at both ends. The "production to evaluation" role shift also offers a candidate explanation for Does AI collaboration drain motivation when workers return to solo tasks?'s finding that intrinsic motivation drops once GenAI absorbs a task's "intrinsically motivating parts" — if production is where engagement lived, shifting it to evaluation is the mechanism, stated here at the role level rather than measured as a psychological outcome.

The excerpt is a research synthesis, not a new study: its "four key reasons" are argued by citing three decades of prior Human Factors literature and a handful of recent GenAI usability studies (on Copilot and other programming tools), not a dataset the authors collected themselves. No effect sizes, task counts, or participant numbers back the claims here. What follows at that strength is a design hypothesis — automation research's remedies are plausible candidates for GenAI tools — rather than a demonstrated fix, which the paper's own second half (its proposed design solutions) takes up directly.