Algorithmic Monoculture

Imagine you apply to twenty companies for an apprenticeship. All twenty pre-screen applications with the same purchased software. Then you don’t get twenty independent chances, but essentially just one — repeated twenty times. This exact pattern is what the term Algorithmic Monoculture refers to: many decision-makers, but only one shared judgment behind them. The word is borrowed from agriculture, where a monoculture means that only a single crop variety grows across vast areas. Such fields yield high harvests, but a single well-suited pest can wipe out the entire crop.

When a single rejection applies everywhere

A society normally relies on decision-makers thinking differently from one another. One HR manager values practical experience, the next focuses on grades, a third on the personal impression made in an interview. Anyone who fails with one has a genuine new chance with the next. This diversity is an invisible safety net. A monoculture cuts right through it.

Researchers call the consequence systemic exclusion. Those affected are not randomly changing people, but always the same ones — namely those the one model in use rates poorly. For the individual, this feels like bad luck. Statistically, it isn’t bad luck at all, but a structural barrier. Whoever falls through the grid falls through it everywhere.

On top of that comes a risk for everyone: everyone depends on the same flaw. If a widely used model has a weakness, that weakness shows up simultaneously in thousands of companies. A single faulty system can thus push an entire market in the same wrong direction. Experts speak of a correlated risk — errors don’t occur scattered about, but bundled together.

Why everyone ends up using the same model

The path into monoculture is rarely planned. Large language models, meaning AI systems that understand and generate text, cost hundreds of millions to develop. Only a handful of corporations can manage that. Everyone else rents access and builds their products on top of it. As a result, countless applications share the same brain.

This is reinforced by economic incentives. Whoever buys the system considered the industry standard never has to justify their choice. If something goes wrong, well, it was the software everyone uses. This safeguard is convenient and pushes companies further toward the same decision.

An important distinction: monoculture does not automatically mean bias. A system can judge in a biased way without being widespread — in which case the damage stays local. And a very fair system can still form a monoculture, because no model is perfect. So the real problem isn’t the quality of one system, but the absence of alternatives alongside it.

From job applications to loan approvals

The effect is most visible in job applications. Many companies use the same programs for pre-selection, often from just a few providers. It looks similar with loans: in Germany, banks rely heavily on scores from a handful of credit bureaus. A poor score then doesn’t just affect one bank, but practically all of them at once.

In tech news, you mostly encounter the term in connection with the market power of a few AI providers. Critics warn that schools, authorities, and newsrooms increasingly write, check, and evaluate using the same model. There is also debate about whether texts online are becoming more similar to one another as millions of people phrase things using the same tool.

Countermeasures exist, but they are inconvenient. Authorities can mandate that important decisions may not be made by automation alone. Companies can deliberately mix different providers or build in random elements so that it isn’t always the same people being filtered out. Every one of these solutions costs efficiency — and that is precisely why monocultures are so persistent.