Until recent­ly, com­pa­nies at least had to make the same mis­takes inde­pen­dent­ly. One orga­ni­za­tion might over­val­ue pres­ti­gious uni­ver­si­ties. Anoth­er might mis­take con­fi­dence for com­pe­tence. A third might qui­et­ly screen out uncon­ven­tion­al careers. Their judg­ments were often flawed. But they were flawed in dif­fer­ent ways.

Now we are build­ing sys­tems that allow orga­ni­za­tions to make the same mis­take together.

Much of the debate around AI asks whether machines can make bet­ter deci­sions than humans. Rea­son­able ques­tion. Pos­si­bly the wrong one. A more con­se­quen­tial ques­tion is what hap­pens when large num­bers of orga­ni­za­tions begin rely­ing on the same sys­tems to decide on their behalf.

A recent study of more than four mil­lion job appli­ca­tions across 156 employ­ers points toward an answer. The researchers describe the emer­gence of an »algo­rith­mic mono­cul­ture«: a sit­u­a­tion in which orga­ni­za­tions increas­ing­ly depend on the same ven­dors, the same mod­els, and ulti­mate­ly the same log­ic for eval­u­at­ing candidates.

The term shifts the focus. Sud­den­ly the issue is not only whether a sys­tem is biased, but what hap­pens when every­one uses it.

From Bias to Monoculture

The usu­al con­cern is bias. This study points some­where else. Even if a sys­tem per­forms rea­son­ably well, what hap­pens when hun­dreds of orga­ni­za­tions use it at once?

For most of mod­ern eco­nom­ic his­to­ry, orga­ni­za­tions devel­oped their own ways of iden­ti­fy­ing tal­ent. They hired dif­fer­ent peo­ple. Reward­ed dif­fer­ent qual­i­ties. Inter­pret­ed poten­tial through dif­fer­ent lens­es. Some approach­es were bet­ter than oth­ers. That is almost beside the point. The vari­a­tion served a purpose.

A can­di­date reject­ed by one employ­er could still suc­ceed else­where because some­body else val­ued dif­fer­ent sig­nals. Dif­fer­ent expe­ri­ences. Dif­fer­ent forms of promise. Mono­cul­tures work differently.

In agri­cul­ture, mono­cul­tures are effi­cient until they are not. They sim­pli­fy pro­duc­tion, cre­ate con­sis­ten­cy, and scale extra­or­di­nar­i­ly well. They also reduce a system’s abil­i­ty to absorb sur­pris­es. Diver­si­ty dis­ap­pears long before the con­se­quences become vis­i­ble. The same pat­tern may be emerg­ing in orga­ni­za­tion­al decision-making.

The Cost of Convergence

As com­pa­nies adopt the same mod­els, they begin shar­ing the same assump­tions about com­pe­tence, poten­tial, and fit. Can­di­dates out­side those assump­tions encounter the same bar­ri­ers repeat­ed­ly. Not because dozens of orga­ni­za­tions arrived at the same con­clu­sion inde­pen­dent­ly. Because they adopt­ed the same way of reach­ing con­clu­sions. Those are not the same thing.

The most inter­est­ing ques­tion raised by AI is not whether a par­tic­u­lar mod­el is fair. It is whether we are grad­u­al­ly stan­dard­iz­ing judg­ment itself. Hir­ing is only the clear­est example.

Orga­ni­za­tions have been mov­ing in this direc­tion for years. They attend the same con­fer­ences, hire the same con­sul­tants, bench­mark against the same suc­cess sto­ries, and increas­ing­ly exper­i­ment with the same AI tools. No sin­gle deci­sion changes much. Tak­en togeth­er, they can pro­duce orga­ni­za­tions that appear dif­fer­ent on the sur­face while becom­ing sur­pris­ing­ly sim­i­lar under­neath. We tend to treat con­sis­ten­cy as a virtue. Some­times it is.

But dis­agree­ment keeps sys­tems flex­i­ble. Vari­a­tion cre­ates alter­na­tives. Inde­pen­dent judg­ment opens paths that stan­dard­ized process­es rarely con­sid­er. These are not inef­fi­cien­cies wait­ing to be opti­mized away. They are part of how com­plex sys­tems learn, adapt, and dis­cov­er pos­si­bil­i­ties that more uni­form sys­tems often miss.

The great­est risk of AI may not be that it makes bad deci­sions. Orga­ni­za­tions have always made bad deci­sions. The greater risk is that thou­sands of orga­ni­za­tions begin mak­ing the same deci­sion, guid­ed by the same assump­tions, at the same time.

We tend to dis­cuss diver­si­ty in terms of peo­ple. The hard­er ques­tion is whether our sys­tems are becom­ing less diverse than the peo­ple inside them. Once orga­ni­za­tions start see­ing the world through the same mod­els, sur­face dif­fer­ences become less inter­est­ing. The real ques­tion is whether any­one is still notic­ing what every­body else overlooked.