Sludge Audit

In the [YOU] on AI Field Guide

The cycle that [YOU] on AI inaugurates takes seriously both the genuine liberation that AI offers and the genuine cognitive costs that unstructured AI use imposes. The sludge audit is the instrument through which these two concerns can be made operational rather than rhetorical. An organization that deploys AI tools without conducting a sludge audit has, by omission, chosen the default of maximum efficiency—and maximum efficiency, in the absence of protective friction, produces workers who go faster and understand less, students who produce more and know less, organizations that ship more code and comprehend less of what they have built.

The audit’s most consequential application is in software development, where the pre-AI workflow contained both categories in close proximity. Dependency management, configuration files, and mechanical boilerplate were sludge: their removal by AI tools is an unqualified gain. Embedded within the same hours were the moments of unexpected error that forced comprehension of system architecture—ten minutes in a four-hour block, indistinguishable from the surrounding tedium, depositing the diagnostic intuition that separates a senior engineer from a junior one. The AI removed both simultaneously, because both presented as friction. The sludge audit requires identifying and preserving the ten minutes while celebrating the removal of the three hours and fifty minutes that surrounded them.

The educational application is equally urgent. The teacher who stops grading essays and starts grading questions—requiring students to produce the five questions that would need to be answered before a worthwhile essay could be written—has conducted an informal sludge audit. The essay production process contained both the sludge of first-draft mechanics and the protective friction of independent engagement with the problem. The question-first structure preserves the protective friction while eliminating the mechanical barrier. The ascending friction that the cycle identifies as AI’s most important cognitive effect is, viewed through the lens of the sludge audit, the relocation of all remaining friction into the protective category—a world in which everything the AI can do is sludge and everything it cannot do is protection.

Origin

The concept builds on Sunstein’s distinction between sludge—friction that serves no beneficial purpose for the person experiencing it—and protective friction. Sunstein and his collaborators developed the sludge concept in the context of behavioral policy: the deliberately difficult subscription cancellation process, the form designed to discourage welfare claims, the bureaucratic procedure maintained by institutional inertia rather than user benefit. The application to AI-mediated work is an extension of the same analytical framework to a new domain where the distinction has become urgent.

The temporal asymmetry that makes the audit necessary is a well-documented feature of human judgment: sludge is immediately recognizable as waste, while protective friction is experienced as sludge in the moment and recognized as valuable only in retrospect. The student who wants the calculator cannot feel the understanding that the long-division struggle is depositing; the understanding reveals its value only later. Any system optimized for immediate user preference will eliminate both kinds of friction, because both feel identical to the person experiencing them in the moment. The audit intervenes at the design level, before the preference is expressed, to classify friction by its developmental function rather than by its experiential quality.

The nudge framework provides the institutional vocabulary for implementing the audit’s conclusions: defaults set to preserve protective friction, choice architectures that require human engagement in high-value categories while automating the rest, structured pauses that create moments for reflection without prohibiting continuation. The audit is the diagnostic step; the nudge is the implementation.

Key Ideas

The two kinds of friction. Sludge is friction that serves the institution that imposes it, or no one at all. Protective friction serves the person experiencing it, building understanding, judgment, or embodied competence that cannot be acquired by any other means. The practical difficulty is that both feel identical to the person in the moment: both are obstacles, both cost time, both feel like impediments to the work. The audit makes the distinction by asking not how the friction feels but what it builds.

The temporal asymmetry. Sludge is immediately recognizable as waste. Protective friction is experienced as sludge in the moment and recognized as valuable only in retrospect, sometimes much later. This asymmetry means that any system optimized for user satisfaction—including every commercial AI product currently available—will eliminate both indiscriminately. The audit is the institutional counterweight to this asymmetry: it makes the retrospective value of protective friction visible at the design stage, before the tool removes it in the name of efficiency.

Category-level implementation. The audit does not require evaluating every instance of friction individually. It requires establishing categories. Implementation is sludge; comprehension is protective. Execution is sludge; diagnosis is protective. The categories, once established, can be applied systematically to configure AI tools that automate sludge categories and require human engagement in protective ones. The configuration is a one-time design investment that produces ongoing developmental returns.

Distributional qualification. The protective friction that experienced practitioners mourn was built on a foundation of access: access to training, mentorship, and the institutional support that transforms struggle into directed development. For practitioners who lacked that access, much of the same friction was not protective but exclusionary—a barrier maintained by structural inequality rather than developmental necessity. A sludge audit calibrated only to the experience of the privileged practitioner tells only half the truth. The design challenge is to preserve friction that is genuinely protective and to eliminate friction that is merely exclusionary, which requires knowing, for each user population, which friction builds and which merely blocks.