In a market where a buyer can’t directly verify a seller’s quality, a signal only separates good from bad if it’s genuinely costly to produce, and differentially so: cheap for someone with the underlying quality to generate, expensive for someone without it to fake. A university degree works as a labour-market signal for this reason. It’s costly enough, in time and effort, that faking it convincingly is harder than actually having the underlying ability the degree is meant to represent.
The theory explains why fluent writing, structured argument, and confident professional register used to carry information about a person’s competence: producing them well required the competence. Generative AI has removed that asymmetry for a wide class of signals. Polished prose and well-organised analysis are now available at near-uniform cost regardless of whether the person producing them understands the underlying material, which collapses what economists call a separating equilibrium into a pooling one. Buyers, rationally, stop trusting the signal and start discounting everyone uniformly.
The response to a broken signal isn’t to produce more of it. It’s to find or build a new one that’s still expensive to fake: a verifiable track record, exposure to the consequences of being wrong, or specific, checkable detail that a general system can’t reconstruct from the public record.