Add Silicon Canal to your Google News Feed. 
The pitch was fascinating. We fed thousands of successful hires into an algorithm, taught it patterns, and watched it filter candidates with surgical precision. You no longer have to make decisions based on intuition. No more unconscious bias from tired recruiters looking through 300 resumes before lunch. AI will make employment fairer.
Except that happened.
A growing body of research suggests that AI-powered recruitment tools are replicating, and in some cases amplifying, the exact biases they are designed to eliminate. And the biggest problem is that companies implementing these systems often don’t realize it’s happening.
The problem with training data that no one wants to talk about
Every machine learning system is a reflection of its inputs. If you train a hiring algorithm on 10 years of hiring experience at companies that have historically promoted white men to leadership positions, the system will learn to favor white men. This should surprise no one. Yet vendor after vendor continues to market these tools as “unbiased” or “objective.”
Amazon made the painful discovery in 2018 when it discovered that its internal recruiting tools were systematically downgrading resumes that included the word “female,” such as “captain of women’s chess club.” The system learned from a decade of hiring patterns in which men dominated technical roles. Although Amazon has discontinued this tool, the underlying logic persists across the industry.
A Bloomberg study published in 2024 found that large-scale linguistic models used in resume screening consistently ranked candidates with names perceived as white and male consistently higher than equally qualified candidates with names perceived as black or female. The models were not told to discriminate. They absorbed it from world texts.
The illusion of “objectivity”
Here’s what actually happens psychologically when companies implement AI recruiting tools. Decision makers let their guard down. They believe the system is neutral, so they stop questioning its output. This is a well-documented cognitive phenomenon called automation bias, where humans defer to decisions made by algorithms even when those decisions are visibly flawed.
A 2023 study from the Center for Gender Research at the University of Cambridge found that HR professionals using AI screening tools were significantly less likely to override recommendations they personally disagreed with compared to overriding recommendations from human colleagues. This algorithm had an authority that human recruiters could not match, simply because it was perceived as mathematical.
This creates a perverted feedback loop. The biased output is left alone, producing biased recruitment data and training the next iteration of the model to be even more biased. This system will self-reinforce faster and at scale, just like human biases.
Video interviews and body language myths
Resume screening is only one page. AI-powered video interviewing platforms used by companies like Unilever, HireVue, and many others claim to evaluate candidates by analyzing facial expressions, tone of voice, and word choice. Premise: Microexpressions and speech patterns reveal traits such as integrity, teamwork, and leadership potential.
Science doesn’t support this. A groundbreaking review by a coalition of AI researchers published in Psychological Science in the Public Interest concludes that the idea of reliably reading emotions from facial movements is at best scientifically debatable, and at worst pseudoscience packaged in a slick dashboard.
Neurodiverse people, people who speak English as a second language, and candidates from cultures with different norms regarding eye contact and expressiveness are unfairly penalized by these systems. Bias is not a bug. It’s built into the methodology itself.
Regulation is coming, but slowly
New York City’s Local Law 144, which went into effect in 2023, requires companies that use automated hiring decision-making tools to conduct annual bias audits. The EU’s AI law, finalized in 2024, classifies AI employment tools as “high-risk” systems, subject to transparency and monitoring requirements. Similar legislative battles are unfolding in Illinois, Maryland and across the Southeast.
However, the introduction of regulations has been delayed for years. A 2024 study by the Society for Human Resource Management found that 79% of employers using AI in hiring implemented the tools before relevant regulations existed in their jurisdiction. Many companies still operate without accountability.
The vendor itself is useless. Most companies refuse to release their models for independent auditing, citing proprietary algorithms. If an audit is conducted, it is usually conducted by a company chosen and paid for by the vendor itself. Conflicts of interest are structural.
What this actually means for job seekers
If you’ve applied for a job at a medium to large company in the past three years, an algorithm will almost certainly make the first decision on your candidacy. You probably didn’t know. There was no disclosure, no explanation, and no appeal process.
For candidates, this creates an asymmetric information space that is nearly impossible to navigate. You can’t optimize a system you can’t see. And the advice circulating online (use specific keywords, reflect your job description, maintain eye contact with the camera) amounts to folk wisdom disguised as strategy.
A more serious issue is the issue of organizational trust. Companies are deploying these tools to demonstrate fairness and modernity. Candidates encounter an opaque and unaccountable system that treats them as a data point. The gap between marketing and reality undermines trust in the entire recruitment process.
unpleasant question
The real question is whether these tools were primarily intended to reduce bias. The economic incentive for AI recruitment tools is efficiency. That means you can process 10,000 applications at a fraction of the cost of human review. Reducing stigma is your brand story. Speed and cost reduction are priorities.
This is important because it determines how companies evaluate tools. If the metric is “time in action” then the system appears to be successful. When the criterion is “Did we build a more equitable workforce?” almost no one checks. There is still surprisingly little research into whether AI adoption leads to better and more equitable outcomes over time.
What we are left with is an industry-wide experiment with live candidates, with limited oversight, debatable science, and a marketing narrative that outweighs the evidence by a decade. Algorithms are not neutral. They never were. And until companies start treating it as a design issue rather than a PR issue, bias will continue to silently worsen at the speed of automation.
Featured image by Markus Winkler on Pexels
