Platformer: the AI productivity paradox

Casey Newton’s January 2026 piece for Platformer examines a pattern that has emerged across multiple surveys: managers report significant productivity gains from AI, while workers at the same organizations report those gains are largely absent from their daily experience.

The most striking data point comes from a METR study of experienced open-source developers. Given access to AI tools, these developers took 19% longer to complete tasks than when working without them. At the same time, they reported feeling 20% more productive. The gap between perceived and actual performance is not a rounding error — it points to something structural about how people evaluate their output when AI is involved.

Newton’s explanation focuses on where the gains actually land. Executives generate substantial output: emails, presentations, reports, strategy documents. AI tools accelerate this kind of work considerably. But the material that results — AI-generated content that looks polished but requires review and correction — often flows downstream to subordinates who must validate and fix it. The efficiency gains at the top create additional work for the people below.

The broader survey data supports this reading. A Section survey found two-thirds of employees saved zero to two hours per week using AI, with 40% saying they would be comfortable abandoning it entirely. A PwC survey found 56% of companies report gaining nothing from AI investments, despite significant spending.

For writers and editorial teams, the implications are specific. AI appears to benefit generative tasks — drafts, summaries, copy variations — more reliably than it benefits tasks that require judgment, synthesis, or source development. When AI tools are adopted primarily to cut costs rather than to improve editorial quality, the productivity math rarely holds up against the projections.

Newton’s practical recommendations are directed at both sides of the gap: managers should base assessments on measurable analytics rather than enthusiasm or self-reporting; workers should build genuine understanding of what AI can and cannot do, both to use it more effectively and to protect against the risks of uncritical dependence.