A Program Reads Your Application

Passage A: Use the Tools, but Check Them

1When people apply for a job, someone has to decide which of them will get an interview. For most of history, that has been a person. Today, many employers use computer programs, including artificial intelligence (AI), to do part of that work. In a 2024 survey by the Society for Human Resource Management, a national organization of people who work in human resources, about one in four organizations said they used AI in their human resources work. Of those, about two in three used it for recruiting, interviewing, or hiring. Employers should be allowed to use these tools, but only under clear rules that are checked.

2The first reason is that people who hire are not as fair as we like to believe. In an experiment published in 2004, the economists Marianne Bertrand and Sendhil Mullainathan answered help-wanted ads in Boston and Chicago with made-up résumés. Some carried names common among white Americans, such as Emily and Greg. Others carried names common among Black Americans, such as Lakisha and Jamal. The names were assigned at random, so neither group was better qualified. Even so, résumés with white-sounding names received 50 percent more calls for interviews. The employers who made those calls were people, not programs. The real choice is not between a biased machine and a fair human being. Both can be unfair.

3There is evidence that a scored tool can do better than a person working alone. In a study published in 2018, the economists Mitchell Hoffman, Lisa Kahn, and Danielle Li looked at 15 companies that began using a job test, scored by computer, to hire for service jobs similar to data entry and call-center work. Managers who often hired people the test had ranked low ended up with workers who left their jobs sooner, on average. The researchers concluded that managers often overruled the test because they were “biased or mistaken.” This test was not AI, but the lesson carries over: a tool that scores every applicant the same way can catch mistakes that people make.

4Supporters of AI should not pretend that these tools are always fair. In 2018, the news service Reuters reported that Amazon had built an experimental program to rate job applicants, and that it marked down résumés that included the word “women’s.” But the company’s own team found the problem, and Amazon gave up the project. A program can be tested again and again with résumés that are the same except for one detail, such as a name. It is much harder to test what goes on inside a hiring manager’s head.

5New York City has already shown what rules can look like. Under Local Law 144, enforced since July 5, 2023, an employer that uses an automated tool to screen applicants for jobs in the city must have it checked for bias each year by an independent auditor, post a summary of the results online, and tell applicants at least ten business days ahead. Other places should adopt rules like these and go further: a person, not a program, should make the final decision. In a Pew Research Center survey published in 2023, 47 percent of adults said AI would be better than humans at treating all job applicants the same way; only 15 percent said it would be worse. With honest testing and a person making the final choice, AI can help make hiring fairer than it has been.

Passage B: Keep People in Charge of Hiring

6When a person turns you down for a job, you can at least imagine asking why. When a program turns you down, often within minutes, you may never learn that a program was involved, or what it looked at. Employers should not let AI decide who is considered for a job and who is screened out. If employers use these tools at all, the limits should be strict, and they should be enforced.

7AI programs learn from the past, and the past is full of bias. In 2018, Reuters reported that Amazon had spent years building a program to rate job applicants from one to five stars. The program learned from ten years of résumés sent to the company, most of them from men, and it taught itself that men were the better candidates. It marked down résumés that included the word “women’s,” as in “women’s chess club captain,” and it marked down graduates of two all-women’s colleges. Amazon gave up the project. The lesson remains: a program trained on yesterday’s hiring decisions will repeat yesterday’s unfairness, and it will do it faster.

8Newer AI shows the same problem. In 2024, researchers at the University of Washington, led by Kyra Wilson and Aylin Caliskan, asked three AI language models to rank more than 550 real résumés for real job descriptions, changing only the names. The models favored names associated with white people 85 percent of the time, and names associated with women only 11 percent of the time. They never preferred a name associated with a Black man over one associated with a white man. The models in the study were not hiring products sold to employers. Still, they are the same kind of technology that employers are now being offered.

9When one program screens applicants for many employers, one flaw can harm a great many people. In 2023, Derek Mobley, a Black man over 40, sued the software company Workday in federal court in California. He says he applied for more than 100 jobs at companies that used Workday’s screening software and was turned down every time. In May 2025, a federal judge allowed his age discrimination claim to go forward as a collective action, which meant that other applicants over 40 could join it. Workday says the case is without merit. As this is written, in the fall of 2026, the court has not decided whether his claims are true.

10Rules on paper are not enough. New York City’s Local Law 144 requires bias audits and notice to applicants, but it does not forbid an employer to use a tool that an audit shows to be unfair. It requires only that the results be posted. In 2024, researchers from Cornell University and other groups checked 391 employers and found only 18 audit reports and 13 notices to applicants. In December 2025, an audit by the New York State Comptroller found that the city agency in charge had received only two complaints in two years. When the agency reviewed 32 companies, it found one problem; the state’s auditors, looking at the same companies, found at least 17 possible problems. In a Pew Research Center survey published in 2023, 71 percent of adults opposed letting AI make the final hiring decision. Until the rules are strong and enforced, people, not programs, should decide who is screened out.