Part of a series “old econ papers”
Michael Spence’s Job Market Signaling starts from an asymmetry: employers can’t directly observe how productive I am, but I know more about myself than they do. I can choose an observable signal, like getting a degree — and the signal doesn’t even need to make me more productive to be useful. It only needs to be cheaper for high-productivity people to acquire.
Education can be useful even if it teaches nothing
Take a toy example: a degree gives me a €50k expected lifetime wage premium, it costs a high-productivity person €20k worth of effort and time, and it costs a low-productivity person €80k. Then for the high type, 50 − 20 > 0, so they get the degree. For the low type, 50 − 80 < 0, so they don’t. Employers observe the degree and correctly infer something about hidden productivity.
The society can spend a lot of resources on an activity whose private value comes partly from proving who you already were. Of course real education can both teach skills and signal them, but how much of each?
A modern estimate says signaling is substantial but not everything
Signaling and Employer Learning with Instruments estimate a private return to education of 7.9%, decomposed into:
- 70% increased productivity
- 30% signaling
AI destroys some signals almost overnight
Now take a signal like:
“This applicant writes an excellent cover letter.”
Before LLMs: good cover letter ≈ writing skill + effort + conscientiousness.
After LLMs: good cover letter ≈ ability to open ChatGPT.
When a signal becomes cheap for every type, it stops separating types, and the market searches for another one.
AI also makes screening cheaper
Signaling exists partly because employers can’t cheaply measure true ability, so what if AI lets them screen candidates much more directly? Instead of inferring “degree → probably good at X,” a company may run a cheap personalized evaluation of X, analyze a candidate’s actual past work, or simulate job-relevant tasks.
AI may change the underlying type too
Suppose the “low type” becomes much more productive with AI. In Generative AI at Work, less experienced and lower-performing customer-support agents benefited more from AI assistance than stronger workers, so AI can compress the productivity differences that the original signal was trying to reveal.
I expect credential churn
My guess is that we are going to see a lot of old signals lose value and new ones appear. When everybody can write a polished application, polish is meaningless; when everybody can vibecode a demo, “I built an app” means less. Maybe the scarce signal moves toward:
- did anybody use it?
- did it survive for a year?
- did another competent person choose to work with you?
- can you repeatedly make good decisions when the problem is underspecified?
Check Why You Need to Read Sartre in the Age of Agentic AI: AI makes more of the surface area of competence cheap, which makes direction, taste and actual outcomes more informative.