AI Lets Novices Impersonate Experts — And Workplaces Reward the Illusion

The author describes a pattern emerging in knowledge work: generative AI lets people produce artifacts that look expert without being expert. Two distinct failure modes are conflated in current research. The first — novices producing senior-grade work in their own field — is well studied. The second, riskier mode is cross-domain generation: people building software, designing data systems, or producing technical artifacts in disciplines they were never trained in, with no ability to evaluate what they ship. A central anecdote describes a non-engineer colleague who spent two months building a data architecture that was wrong from day one, defended by managers invested in the appearance of momentum.

The piece coins this output-competence decoupling: work quality used to be a reliable signal of the producer’s skill, because doing the work was how judgment was acquired. AI severs that link, turning the worker into a conduit who routes output they cannot assess. Studies cited (Stanford on model sycophancy, NBER on support agents, HBS on consulting) show LLMs boost novice productivity disproportionately while flattering users into overconfidence — a combination that produces convincing artifacts no one in the loop can vet.

The second-order effect is institutional. Documents bloat because generation is free but reading isn’t; checkpoints drown in synthetic prose. Agentic systems are designed around the premise that humans are the bottleneck, but the human was the only party with skin in the game and the only mechanism for the system catching itself. The pipeline of future experts thins from both ends as the work that taught judgment gets handed to the tool.