The largest empirical study of algorithmic hiring yet ran the numbers this week, and the headline finding is quieter than it should be. Over 90% of U.S. employers now screen applicants with hiring algorithms, and many of them buy from the same handful of vendors. Researchers at Stanford, Chapman, and Northeastern pushed 3.4 million real applicants — four million applications across 156 employers and 11 sectors — through a single vendor's system, and found measurable adverse impact against Black and Asian applicants. The catch: it was visible only when they disaggregated the results position by position. In aggregate the screen looked fine.
The paper reaches for the right word. Not bias — *monoculture*. The problem isn't that the algorithm is wrong; a wrong algorithm is a tractable thing, you fix it or you sue it. The problem is that it's the *same*. One vendor's judgment, replicated across every employer that licenses it, so a candidate it filters out isn't rejected by a company — they're rejected everywhere that runs the same gate, at once, for the same reason, with no second opinion anywhere in the system.
We conduct the largest empirical study of algorithmic hiring... we test whether this algorithmic monoculture bottlenecks job opportunities.Algorithmic Monocultures in Hiring
And it isn't only hiring. The BBC spent the same week documenting how social feeds stopped being social — how the stream that once surfaced the people you actually know now surfaces whatever's trending, fads instead of friends. Same shape: a shared optimizer, tuned for one signal, flattening a whole domain into a single homogenized output that converges on the same thing for everyone. The social graph was variety. The engagement model is a monocrop.
A blogger gave the engine a name this spring, and it's a good one. He calls it dopamine fracking:
The act of pumping immense, disproportionate resources — money, crowdsourced math, analytics, optimization, min-maxing — into a previously casual or complex, layered activity to forcefully extract and squeeze out the purest, most concentrated dopamine hit, with no regard for anything except dopamine.igerman.cc
Fracking is exactly the metaphor: enormous short-term yield, long-term depletion of the thing you drilled into. Monoculture works the same way in a field. You plant the one crop with the best yield, everywhere, and for a few seasons it's a triumph — until the blight arrives and discovers that every plant it meets is identical.
For two years the worry about AI has been that it will be *wrong* — that it hallucinates, libels, misfires. This is the subtler failure, and I'd argue the worse one: AI being uniformly, confidently the *same*. A thousand mediocre human screeners disagree with each other, and that disagreement is slack in the system — a bad call in one office is survivable because the next office calls it differently. Replace all thousand with one good-enough model and you've removed the slack. The errors stop averaging out. They correlate.
A thousand mediocre screeners disagree with each other. One good-enough model does not — and that's the whole risk.
Which means the fix isn't a better model. A better model run by everyone has the same shape of failure as a worse one; you've just raised the floor and kept the correlation. The fix is variety — more than one gate, more than one optimizer, enough independence in the system that the parts can still disagree. We spent the decade optimizing for the single best answer. The thing actually worth protecting is the second one.