Urgent.News

What's breaking now, across thousands of outlets.

AI

Cómo superar la 'monocultura algorítmica' que nos rechaza en todas las empresas

La inteligencia artificial prometía hacer más objetiva la selección de personal. Pero si compañías distintas utilizan sistemas que miran a los candidatos de manera parecida, un error deja de ser el problema de una empresa: puede convertirse en una barrera que acompaña al trabajador por todo el mercado laboral. El uso de herramientas similares puede erosionar una de las ventajas del mercado…

Cómo superar la 'monocultura algorítmica' que nos rechaza en todas las empresas

A demand presented against Meta in July highlighted a concern from a group of workers who claim that internal measurement and evaluation systems could convert circumstances like a medical leave, family leave or disability into signals of lower performance. Though the accusations still need to be proven, the case illustrates how an algorithm does not necessarily have to explicitly use a protected characteristic to harm the person possessing it; it is enough to penalize a variable that functions as a substitute for it.

Now, a new line of research expands the problem: what happens when many companies use similar ways to measure and dismiss candidates? Researchers from Stanford University call this phenomenon algorithmic monoculture - similar systems, data and criteria can cause the same mistake to repeat and accompany the candidate from one company to another.

So far, most of the debate about AI and hiring questioned whether a specific tool discriminated within a specific company. Stanford raises a different issue: what happens across the entire labor market when many companies use systems that make decisions similarly. The researchers analyzed four million applications from over three million people processed with technology from a single provider, and found that 4% were rejected across ten different positions, and around 10% failed to pass any of the processes.

This pattern, called systemic rejection, occurs because the results repeat more than would be expected if each company evaluated candidates truly independently. Algorithmic monoculture arises when different companies end up looking at candidates almost with the same eyes. They don't need to explicitly use the same program; it's enough for their systems to be trained on similar data, value the same skills, or share a similar idea of what makes for a good profile.

The result is that a person rejected by one company may be rejected by many others for very similar reasons. What once was a solitary opinion begins to become a common market criterion. The novelty of algorithmic monoculture is not discovering that an algorithm can be wrong or discriminate; it's discovering what can happen when the same type of error repeats across different companies.

Until now, the labor market offered job seekers a second chance: being rejected by one company didn't mean being rejected by the others. If many companies use systems that look at talent similarly, that second chance may start to disappear. The study also detected racial differences: 25.87% of applications from Black people and 14.74% from Asian people corresponded to positions where these groups had worse results.

The important thing is that the system didn't receive data on race, gender or age. It evaluated candidates through games. This means that an algorithm can generate inequalities even without directly using protected characteristics, because other seemingly neutral traits can produce similar discriminatory effects indirectly. This possibility was anticipated five years ago by a study published in Proceedings of the National Academy of Sciences.

It concluded that even a fairly accurate algorithm can worsen market functioning if everyone depends on it. The reason is that different companies may end up committing the same mistakes and dismissing the same people. The search for a job changes as well with algorithmic monoculture. Sending more resumes used to increase chances because each company could value the candidate differently.

But if many companies use similar tools, filters or criteria, a new application may be evaluated almost the same way as the previous one. In other words, changing companies may not always mean facing a truly different evaluation. That's why ten rejections can have two very different explanations: either ten companies independently reached the same conclusion, or they're all using systems that tend to value the candidate similarly.

The problem is that the person usually doesn't know which tools have been used or why they were rejected. Stanford researchers warn that in this scenario, some candidates may need to submit themselves to many more offers to increase the chances of an application getting past the automatic filter and be reviewed by a human. This helps to understand the distrust of candidates as well.

A Gartner survey of 2,918 people found that only 26% trusted that AI could evaluate them fairly; 32% feared it would harm their candidacy, and one in four doubted they would trust companies that use AI to evaluate candidates. Putting someone at the end of the process does not guarantee true supervision. A University of Washington experiment with 528 participants showed that human supervision does not always correct machine biases: when evaluators received biased algorithmic recommendations about equally qualified candidates, their own decisions tended to incline in the same direction.

In cases of higher bias, they followed around 90% of the recommendations. The question is not choosing between people or algorithms, but deciding what we want those systems to measure and how much room we leave to discover talent differently. AI can reduce human arbitrariness, but it can also repeat at a large scale the same criteria and mistakes. Korn Ferry warns that it will become increasingly important.

Written by urgent.news from Expansion ES's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at expansion.com →

More in AI

RBAC for AI: Governing the AI Control Plane

Giving someone access to the AI control plane is not the same as deciding which parts of it they are allowed to control. The previous post described virtual keys as the runtime representation of a…

More from Tuesday 18 August →