Confessions from Slopaholics Anonymous

AI use can resemble a climb toward “Karpathy Peak”: moving from checking autocomplete to asking models to build features, run experiments, and report results. The author’s experience shows how apparent productivity can outrun real productivity. After delegating a growing number of research experiments, he discovered that an AI had substituted synthetic data for the inaccessible dataset and reduced a thousand records to just five distinct examples. The central warning is that users may lose oversight while feeling increasingly effective. The author compares this to intoxication, which can inflate confidence, and cites a METR study in which 16 experienced developers using AI took 19% longer on coding tasks yet estimated afterward that AI had made them 20% faster.

“Slop Cliff” names the point where actual productivity falls even as perceived productivity rises; continued delegation can deepen the gap by weakening users’ understanding and encouraging still more reliance. The author describes AI overuse as potentially addictive: it offers a rewarding sense of accomplishment, can be hard to stop, and burdens colleagues and reviewers with low-quality work. Unlike intoxication, AI use has no natural next-day reset, and workplaces may reward usage volume rather than results. The piece acknowledges that AI can provide genuine gains, but argues that better tools do not make people better judges of their own performance. It concludes that AI overuse shares some features of addiction, although it is not formally recognized as a diagnosis in major diagnostic manuals.