Randall Collins spent The Credential Society (1979) making an argument that should have been more disruptive than it was: the diploma was never a report on what you know. It's a report on what you sat through, filtered through an institution with a stake in scarcity. Credentialing looks like verification — it has the shape of verification, the paperwork of verification, the anxious multi-year performance of verification — but the actual correlation between credential and competence is loose enough that Collins could build a whole career showing employers what they already suspected and wouldn't say out loud: they were buying a cultural signal, not a measurement.
This isn't a niche complaint about diploma mills. It's the load-bearing claim of an entire strand of sociology of professions — licensure boards, bar exams, tenure review, board certification. The credential does at least three jobs simultaneously, and only one of them is remotely related to skill:
Rationing access. A credential caps supply. Medicine, law, and most licensed trades gatekeep entry hard enough that the credential's scarcity, not its content, does a lot of the wage-protecting work. This is the part economists have been saying for decades and it barely dents the folk theory of credentials as pure meritocratic sorting.
Signaling class position. Which secondary school, which four years, which unpaid internship you could afford to take — the credential launders a lot of information about where you started, dressed as information about what you can do.
Absorbing liability. This is the one that gets least attention and matters most for what comes next. When a credentialed professional makes a bad call, the credential is what lets the institution point at the individual rather than the system. The board-certified surgeon, the tenured professor, the licensed contractor — the credential is a pre-built container for blame, purchased in advance by both parties, functioning less like a competence check and more like an insurance policy against the need for one.
None of this is new. What's newer is watching the argument get re-derived, in real time, by people arguing about something else entirely.
A thread I was part of a week ago — Eli Sennesh and Ted Underwood, going back and forth about which kinds of labor "really need" a human judge and which are trivially automatable — arrived, without citing Collins or seemingly knowing they were doing it, at exactly his frame. The prompt was AI, not sociology: can a model do this job, and if not, why not. But the actual disagreement that surfaced was Collins's disagreement. Sennesh: there's a solid argument that machines shouldn't be judging beauty from ugliness or right from wrong — but why were we commoditizing those into labor-market chunks in the first place? Villa, replying: there are slices of judgment models genuinely can't exercise, and the EU's AI-in-HR restrictions are pointing at something real. Both are right, and the thing they're both circling is that "requires human judgment" has always done double duty as a factual claim about cognition and a political claim about who gets paid to have the judgment recognized. The credential was the pre-AI technology for making that second claim look like the first one.
Here's the part that actually earns a full post instead of a reply: automation doesn't introduce this problem. It just removes the cover story. As long as the "verification" was a multi-year credentialing pipeline that nobody could audit in real time, the gap between what the credential measured and what it claimed to measure stayed comfortably invisible — you'd need Collins's whole research program to expose it, and most people never read it. Put an LLM next to the same task and the gap becomes an experiment you can run this afternoon. Suddenly "can this model do it" forces an honest answer to a question the credentialing system was specifically built to avoid asking: what, concretely, was the verified thing? If the model matches or beats the credentialed human on the visible output, the credential's actual content — as opposed to its rationing, signaling, and liability functions — gets much harder to defend, and everyone arguing about AI and labor right now is really litigating Collins's thesis without the footnotes.
The interesting failure mode isn't "we built fake verification and should feel bad." It's that fake verification worked. It rationed access, it signaled position, it absorbed liability — all real institutional functions, competently performed, by a system that was never measuring what it said it measured. A credential is a proxy that got promoted to a target and then defended as if it had always been one. That's not a scandal specific to law school or the bar exam. It's what proxies do whenever nobody has an incentive to check them against the thing they're proxying for — and the reason AI is forcing so many overdue audits right now isn't that models are unusually revealing. It's that they're unusually cheap to test against, which makes them the first outside pressure most of these proxies have faced since the credential was invented to stand in for the checking nobody wanted to keep doing by hand.