AI Hiring.
If you've ever been part of a busy hiring cycle, you know how those little tasks can take over your whole day. Between chasing updates, coordinating calendars, and keeping everyone informed, there's plenty of work that people would happily hand over to technology.
And at first glance, technology doesn't disappoint.
Recruiting software can now write job posts, organize applications, do early interviews, and send back a ranked shortlist before the hiring team even looks at each candidate. Each step gives you a simple number to work with.
The harder number to find is how many capable people were mistakenly filtered out along the way.
And it's easy to see why people are drawn to it. In SHRM's 2025 survey, over half of American HR professionals said their companies use AI for hiring. Among those using it, 89% said it saved time or improved efficiency. Processing high volumes of applications is difficult work, and there is little virtue in making people do repetitive tasks that software can handle.
A hiring process is vulnerable if it relies mostly on decision speed as evidence. Short screening times barely tell you anything about how relevant someone is, how consistent the scores are, or the candidates who never make it to an interview.
The quiet inheritance of flawed judgment
Amazon once tried to teach an AI to spot good candidates by feeding it loads of old resumes.
It seemed like a convenient shortcut until the engineers took a closer look at the rankings. As Reuters reported in 2018, the system had picked up gender bias during training and started favoring certain profiles over the rest.
The team tried correcting the obvious signs of bias, but there was no guarantee that the software wouldn't find other ways to make the same distinctions. Eventually, the company abandoned the experiment. Recruiters had seen its recommendations, although they had never relied on them alone.
What makes this particularly troubling is how easily it could have gone unnoticed. It learned to mimic past trends, not realizing if they reflected true talent or just old hiring biases. Those assumptions were now finding their way into scores that appeared objective.
Something similar happened in AI-powered video interviewing, although the problem was different. An interview platform had built facial analysis into its assessments, using visual cues alongside candidates' responses. The technology was sophisticated, but as SHRM reported in 2021, the company's own research eventually found that this added almost no value in predicting job performance beyond what language-based analysis could already provide.
The provider subsequently removed visual analysis from its new models. An external audit of one assessment had reported favorable fairness findings, so the decision wasn't proof that the system was discriminatory. It did, however, raise a fairly basic question. If a characteristic tells us so little about someone's ability to do a job, why was it being measured in the first place?
Of course, human interviewers have biases too. So why should we expect machines to be any different?
Think about a self-driving car involved in an accident. We wouldn't dismiss the incident simply because human drivers cause accidents as well. We would want to know whether the system was properly tested, whether the accident could have been prevented, and whether the same fault could put others at risk.
The same applies to hiring. A human interviewer can make an unfair judgment, but when that judgment becomes part of an automated system, it can be repeated across thousands of applications without anyone noticing the pattern. The whole point of introducing technology is to improve existing methods, not simply reproduce their weaknesses on a much larger scale. That is why AI hiring systems need to demonstrate their reliability through testing and independent evaluation, rather than have their errors excused as another form of human fallibility.
These examples highlight various shortcomings in automated hiring. Sometimes a model picks up the wrong lesson from past choices or focuses on the wrong thing, even if it's technically advanced. In both cases, more applicants mean a wider impact of the error.
The limits of certainty
One big issue with AI hiring systems is that it's tough to tell when something's off. If no one takes a good look at how the system runs or checks hiring data from the past few years, certain problems could fly under the radar. On their own, the decisions might look just normal, even if they're part of a bigger pattern of unfair treatment.
An automated interviewer could ask applicants questions of varying difficulty and then compare their scores as if the tests were all the same.
Screening tools can be biased towards known companies, missing skilled candidates with different experiences.
To the applicant, the decision might seem like a routine rejection with no insight into skill evaluation. The process might still seem consistent to the employer. But consistency is a weak defense when the rule being applied is wrong.
Familiar routes such as data-security certification, testing, pilots, and supplier references often provide assurance to organizations. These checks are valuable yet target particular questions. Having a security certificate proves that important security controls are being handled properly. An AI governance standard
can measure an organization's AI management.
Neither one by itself proves that an assessment score actually shows someone's ability for a certain job. Even if an AI does well on a general benchmark, that doesn't prove it tests the skills that matter in real hiring situations.
That is why hiring requires its own quality checks.
What we choose to call progress
None of this requires abandoning automation. It requires treating assessment quality as a result to be demonstrated rather than an assumption tucked inside a faster workflow. This is where the Institute of Hiring Sciences (IHS) aims to contribute by developing shared standards that define what a sound assessment should measure, how its reliability and fairness should be tested, and what evidence should support its results. Independent audits against these standards would allow buyers to compare systems based on the quality of their assessments rather than relying primarily on promises about speed and efficiency.
