Articles

AI tools also seem to offer something particularly appealing in employment decisions: objectivity. An algorithm, after all, does not have instincts, personal preferences, or unconscious bias in the same way humans do. But that does not necessarily make its decisions neutral. 

A study from Princeton and the University of Chicago put this idea to the test. Researchers used large language models (LLMs), including ChatGPT, Claude, and Gemini, in hiring simulations to examine how biases can develop. As the models received feedback about fictional employees’ performance, they began creating their own stereotypes about fictional ethnic groups. 

For example, after a model was told a member of one ethnic group performed poorly as a doctor, it began avoiding other people from that same group for doctor positions and instead hired them for jobs like janitors. Overall, the AI models showed significantly more occupational segregation than humans, with segregation scores about 65% higher on average.

That matters because AI does not need to be intentionally programmed to discriminate. LLMs are designed to identify patterns and make generalizations from limited data, so as they learn from outcomes, they may also learn patterns we never intended them to learn.

And this concern is not limited to a laboratory. In a 2026 study, Stanford researchers examined real-world data from more than four million job applications across 150 employers for evidence of adverse impact—whether a seemingly neutral practice disadvantages a group. They found that 26% of Black applicants and 15% of Asian applicants had applied to positions where the AI system disadvantaged their racial group. Researchers estimated that roughly 40,000 applications may have advanced in the hiring process if selection rates had been equal across groups.

How adverse impact is measured matters. When the researchers combined results across employers and positions, the system appeared to show no adverse impact. But when they analyzed individual positions, they uncovered the differences in selection rates. An AI vendor’s overall bias analysis may therefore miss disparities within specific positions. Employers should understand how a tool has been tested for bias and monitor whether disparities emerge within their own positions.

The researchers also identified the potential for what they called “algorithmic monoculture.” Job seekers who applied to multiple positions screened by the same AI hiring platform were more likely to be rejected from all of them than if companies made decisions independently from one another. This can happen when employers rely on a common vendor whose models evaluate applicants similarly. In other words, when a single hiring vendor is used by many companies, some applicants may be shut out of jobs.

The repercussions of potential AI bias are beginning to play out in the courts. 

There is the closely watched class action lawsuit against Workday, which challenges whether its AI-powered hiring tools discriminate against applicants based on race, age, and disability. The plaintiff, a Black man over 40 with a disability, claims he was rejected from more than 100 roles, sometimes within minutes of applying, despite being qualified. 

The legal concerns around AI also extend beyond hiring. A federal lawsuit against Meta claims that AI-assisted systems used in the company’s layoff process disproportionately targeted employees on protected medical, parental, or disability leave. According to the lawsuit, these systems evaluated workers’ productivity levels without adequately accounting for time spent on protected leave. This allegedly contributed to their layoff scores and potentially violated several federal and state employment laws. 

These cases are still developing, but they raise an important question for employers: If an AI system helps make an employment decision, who is responsible for making sure that decision is fair?

AI is not inherently bad. It can be incredibly valuable for HR professionals. But it’s important to keep in mind AI’s limitations. AI tools learn from historical data, so if that data reflects existing disparities, AI may copy or even amplify those patterns. AI may also rely on factors that seem neutral but are still connected to race, sex, age, or other protected characteristics. Both can potentially lead to discriminatory outcomes.

Whether AI is being used for recruiting, layoffs, or other employment decisions, HR's responsibility doesn’t disappear because technology is involved. Practical steps employers can take include:

• Understand how AI tools are being used. Know what employment decisions they influence, what factors and data they consider, and how feedback or past outcomes may affect future recommendations.

• Continually monitor outcomes. Regardless of a vendor’s overall assessment, look at what the tool is actually producing within your organization and by individual position. 

• Maintain meaningful human oversight. Establish who owns the employment decision and document how the tool is used in the decision-making process. 

AI may make employment decisions faster and more efficient. But automation doesn’t eliminate accountability for the decisions being made.