The cycle launched by [YOU] on AI documents the twenty-fold productivity multiplier that AI tools deliver to engineering teams, and it does so honestly—reporting both the extraordinary gain and the questions the gain raises about what develops in practitioners who no longer perform the work the tool handles. The optimization trap is the conceptual frame that connects these two observations. The metric went up. The thing the metric does not measure may be going down. The optimization is real. The question is whether it is complete.
Reich’s analysis of the Berkeley study that the cycle discusses—the eight-month ethnographic study of AI’s effects on a 200-person technology company—provides the most precise illustration. By every productivity metric, the AI intervention was a success. Work seeped into previously protected pauses. Multitasking became the norm. Workers reported a persistent low-grade cognitive overload that accumulated over time. The productivity gains were real, but they came at a cost that the productivity metric was structurally incapable of registering. This is the optimization trap in its most characteristic form: the metric goes up and the cost is borne by the workers, their families, the quality of their attention and the depth of their understanding—none of which appears on the dashboard.
The concept originates in Reich’s co-authored System Error (2021), where it was applied to the social media era’s optimization for engagement. But its roots reach deeper, to the foundational insight that the choice of optimization target is itself a political decision. When Karl Marx observed that the factory optimized for the commodity and produced, as its unacknowledged byproduct, the alienated worker, he was identifying an early version of the trap: the metric—output per hour—rose while the cost—the degradation of the worker as a full human being—went unmeasured and therefore uncounted. The vocabulary changes across eras; the structure does not.
In the AI context, the optimization trap connects to what Reich identifies as the distributional question at the heart of the transition. When productivity increases by a factor of twenty, the question of who captures the gain is a question about the distribution of economic power. If the gain flows to the company as higher margins, it enriches shareholders. If it flows to workers as higher compensation or shorter hours, it improves lives. If it flows to the public as lower prices, it produces a genuine public good. The technology itself does not determine the distribution; the institutional arrangements do. And at present those arrangements are designed to direct productivity gains toward capital. The optimization trap produces the gain. The institutional context determines who bears the cost and who captures the benefit.
Metric Substitution. The trap begins when a measurable proxy—engagement, productivity, conversion rate, session length—is optimized in place of the underlying value it was meant to approximate. Engagement was once a proxy for the value users derived from a platform. It became the optimization target. The optimization maximized engagement while degrading the value—producing platforms that were more compulsive and less nourishing. Productivity is a proxy for the value of work. When AI maximizes productivity, it may be degrading the quality of professional development that productive struggle provides.
The Invisible Cost. The optimization trap is most dangerous when the cost is borne by something the metric cannot register: the practitioner’s developing judgment, the quality of attention in daily life, the cooperative skill built through the friction of working with other minds, the depth of understanding that accumulates through encounter with resistant material. These costs are real and consequential; they are simply not visible on the dashboard. The optimization declares success. The cost accumulates unseen.
The Political Decision Disguised as Engineering. The most consequential feature of the optimization trap is that it conceals a political decision as a technical one. Choosing to optimize for productivity is choosing that productivity matters more than the qualities productivity does not measure. This choice is made by corporate actors without democratic input, imposed at scale on millions of workers who had no voice in the decision, and legitimized by the language of neutral methodology. AI governance frameworks, Reich argues, must subject the choice of optimization targets to democratic deliberation rather than leaving them to corporate strategy.