LLMs are exceptionally good at mimicking the surface-level appearance of quality knowledge work — polished reports, convincing code, thorough-looking reviews — without necessarily delivering the underlying substance. Workers, rationally optimizing for proxy measures of quality, increasingly delegate output to AI. LLMs themselves are trained on similar proxy signals (RLHF approval, corpus likelihood) rather than truth or usefulness. The result is a self-reinforcing loop: a simulacrum of work that looks right but may not be, governed by Goodhart's Law — when a measure becomes a target, it ceases to be a good measure.
Sort: