- Artificial intelligence is gaining ground in hiring: an estimated 48% of hiring managers use AI tools to screen résumés and evaluate job applications.
- An IDB experiment with 277 recruiters in Ecuador, each of whom evaluated ten job openings, reveals important differences: human decisions are more variable, while AI decisions tend to follow more consistent patterns.
- Greater consistency, however, does not guarantee fairer decisions: AI can also reproduce social biases. In fact, when candidates are virtually identical, random selection may produce more equitable outcomes.
Imagine two recruiters evaluating the same candidates for a job. One is a human resources professional with years of experience; the other is an artificial intelligence (AI) model. Would they choose the same candidate? You might expect so if both are trying to identify the person best suited for the role.
Yet a 2025 IDB study based on an experiment conducted in Quito, Ecuador, reveals something surprising: human recruiters’ decisions are so variable that, in some cases, they resemble random selection more closely than the decisions of an algorithm designed to mimic human behavior. What does this tell us about how people make decisions—and about the role AI should play in these processes?
From Human Judgment to AI in Hiring
Selection decisions are part of everyday life, shaping everything from who gets into university to who gets a job.
Until recently, these decisions relied largely on human judgment. Today, they are increasingly influenced by AI systems—and, in some cases, made by AI alone.
This shift is especially visible in the labor market: an estimated 48% of hiring managers already use AI tools to evaluate candidates, and their use is expected to keep growing in the coming years.
As AI plays a growing role in determining access to employment, it is essential to understand how it makes decisions and what its use means for both the efficiency and fairness of hiring processes.
How Do Human and AI Decisions Compare?
To investigate these differences, a team of IDB economists adapted the Turing test, originally devised to determine whether a machine can imitate human behavior well enough to be indistinguishable from a person.
Here, the aim was not to determine whether humans or AI were better at selecting candidates, but to assess how closely AI decisions resembled those of human recruiters and what might explain the differences between them.
The experiment involved 277 professional recruiters in Ecuador and four groups of AI agents created with ChatGPT, each with the same number of members as the group of human recruiters. Before the experiment began, the researchers verified that the AI agents understood the task, the human traits they were expected to mimic, and the instructions for responding.
Each recruiter evaluated ten job openings. For each one, they had to choose between two candidates with very similar profiles in terms of education, experience, and age. The only difference between them was a randomly assigned personal characteristic, such as gender, sexual orientation, or Venezuelan migrant status.
In principle, these characteristics should not influence a hiring decision.
More Consistent Decisions—but Are They Fairer?
The study revealed a fundamental difference: human recruiters make less systematic decisions than AI agents, which tend to follow more consistent patterns.
This creates a paradox: the more systematic AI becomes, the further it moves from the way people actually make decisions. Yet the more closely it is designed to reproduce human behavior, the more variable its decisions become—and the more it risks losing one of its main advantages: consistency.
The experiment also showed that, when two candidates are virtually identical, random selection may produce more equitable outcomes than a decision based on systematic criteria.
In other words, when candidates’ profiles do not clearly indicate who is better suited for the job, applying criteria more systematically does not guarantee a fairer decision.
A Question of Bias
These differences between human and AI decisions become especially important when personal characteristics that should not affect hiring come into play, such as gender, sexual orientation, or migration status.
Although AI applies criteria more systematically, it can still reproduce the prejudices and biases present in society. Mistaking consistency for genuine impartiality can lead to excessive confidence in the objectivity of its decisions.
So rather than asking whether AI can imitate human decisions, the more pressing question is how far it reproduces—or even reinforces—existing biases, and how to prevent those biases from translating into discrimination in access to employment.