How to Use AI in Hiring Without Adding Bias
AI can make your hiring less biased, or a great deal more. The technology is not what decides which. How you point it is. Where bias creeps in, the rules that keep it out, and the law you now hire under.
By Umair Ali · ·
Use AI in hiring and the same screening system can strip discrimination out of your process or quietly scale it up, and which one you get comes down to choices you make before the first candidate ever applies.
That is the part most of the coverage misses. The headlines fall into two camps, AI is dangerously biased, or AI is finally objective, and both are wrong in the same way. They treat "AI" as one thing with a fixed effect. It is not. The same kind of system can strip bias out of a process or bake it in, depending on a handful of design and deployment choices that are yours to make. Get them right and AI does the consistent, tireless legwork a fair process needs. Get them wrong and it launders old discrimination into a number that looks neutral.
This guide is for the employer trying to use AI in hiring without becoming the cautionary tale. It walks through the specific places discrimination enters an automated process, the design choices that shut each one down, and what regulators in your jurisdiction now expect of you. It is written for a small team with no compliance department, and none of it is legal advice, so confirm the specifics for where you operate.
Drop the "biased or objective" question
Throw out the question everyone opens with: "is AI biased?" It is the wrong question, and it leads to two bad answers, ban it or trust it, both of which skip the thing that matters.
The useful question is narrower. What job are you giving the AI, and on what information? An AI told to find people who resemble your last ten hires will faithfully reproduce whoever you happened to hire before, bias and all. An AI told to score how confident someone looks on video will punish accents, disability, and nerves. An AI told to check whether a candidate can back up the claims on their own resume, then hand a human the evidence, is doing something a fair process wants done. Same category of tool. Opposite effect. The design is the whole story.
Where AI adds bias
Bias does not wander into AI hiring at random. It comes through a short list of known doors, and someone chose to open each one. These are the main ones.
Training on your own past hires. This is the classic. Amazon scrapped an internal recruiting tool after it taught itself to mark down resumes that mentioned women, because it had learned from a decade of mostly male engineering hires. Any system built to find candidates who look like your previous good candidates will copy the past, and the past is where the discrimination sits.
Scoring face, voice, and accent. Video-interview tools that analyze expression and speech have been found to rate people lower for their accents, their facial expressions, even background noise. That falls hardest on non-native speakers, disabled candidates, and neurodivergent ones. None of it measures whether a person can do the job.
Auto-rejecting behind a black box. Plenty of tools output "recommend" or "do not recommend" with no reason and no human review. When one widely used vendor's model does that across hundreds of employers, the same qualified person can be filtered out everywhere they apply. A large Stanford study called this systemic rejection. If you cannot see why someone was cut, you cannot tell whether it was fair.
Rubber-stamping the machine. This one is quieter and well evidenced. A University of Washington study found that when people are shown a biased AI recommendation, they tend to follow it, and their own choices drift to match the machine's, even when the bias is plain to see. So a tool that only advises can still bend a whole process, if the human treats its ranking as the answer rather than an input.
Hidden proxies. Strip out names and photos and a model can still rebuild protected traits from proxies: a postal code, a school, an employment gap, the name of a club. Removing the obvious signal does not remove the correlation.
The rules that keep bias out
You can shut every one of those doors. None of it means giving up AI, and none of it needs a data-science team. It needs discipline about how the tool is used.
Point it at ability, not identity or history. The AI should weigh what a person can do and prove, not who they resemble or where they are from. A tool that tests claims a candidate makes about their own work stands on safe ground. A tool trained to match your past hires does not.
Keep a human making the decision. Use AI to gather and to check. Do not use it to decide. The final yes or no belongs to a person who can explain it and answer for it. That is no longer just good practice. It is close to a legal requirement in the making.
Refuse face, voice, and accent scoring. If a tool grades how someone looks or sounds on video, do not buy it. There is no version of that which measures competence, and it is the most reliably discriminatory thing in the whole category.
Demand an evidence trail. Every recommendation should arrive with a reason you can read: this claim, checked against this answer. Turn down black-box scores. If a vendor cannot tell you why a candidate landed where they did, you cannot defend the decision, and you may soon have to.
Watch your own automation bias. Treat the AI's output as one input, not the verdict. Force yourself to look at the evidence behind a ranking before you accept it, because the research says that otherwise you will drift toward whatever the machine said.
Know the law where you hire. This is moving quickly, so keep current. As things stand: New York City requires yearly bias audits and public reporting for automated hiring tools. California finalized rules in 2025 on how existing discrimination law applies to them. Colorado passed a broad AI act with duties for both the makers and the users of these systems; it has since been revised, with its requirements now landing in 2027. In the Workday litigation, a court let stand the claim that an AI vendor's tool can act as an "agent" of the employer, which shows where liability is heading. The European Union's AI Act treats hiring as high-risk and regulates it as such. Wherever you are, assume you are responsible for what your tools do, because the law is settling on exactly that.
Where a tool like ours fits
Let me be straight about where I am writing from. BestHire is one of these AI hiring tools. Everything above applies to it too, and you should hold it to the same standard as any other.
Here is the honest account of how it is built against that standard. It interviews every applicant who fits the role and tests the specific claims on their resume, so it is judging what people can back up, not who they resemble or where they studied. It reads only what a candidate says in the interview, never their face, voice, tone, or accent, so the entire video-scoring failure mode is off the table by design, and it collects consent and holds to GDPR. It never auto-rejects anyone. It gives the founder a ranked shortlist with the evidence attached, the exact resume line and the interview answer behind each verdict, so there is no black box, and a person makes every call.
Now the part a careful reader should insist on. None of that makes it bias-free, and no honest vendor can say their tool is. Automation bias applies to us like everyone else. Any tool that ranks candidates can steer a person who stops reading the evidence and just takes the order. Our answer is to hand over evidence instead of a verdict, and to keep you deciding, but that is a mitigation you have to use, not a cure we can promise. A model can still carry bias no one has caught yet. The design lowers the known risks. It does not erase them, and staying fair is work that never fully ends. A tool that tells you otherwise is the one to worry about.
This one is not abstract for me. I once looked at another AI hiring tool that scored candidates on their facial expressions, their background, even their voice, and folded all of it into the ranking. Picture the person on the other end. They record a video assuming what counts is what they say, while the tool quietly grades how they look and sound, accent included, none of which tells you whether they can do the job. That is when I decided BestHire would read the transcript and nothing else. Ordinary hiring guidance already warns against judging someone on their voice or their accent, and a tool that does it at scale is just bias with a dashboard.
Does AI make hiring fairer
The truthful answer is that it depends on you, and that is not a dodge. Used to replace human judgment with an opaque score, AI concentrates bias and hides it. Used to do consistent verification that a responsible human then weighs, it can cut the bias that slips into tired, rushed, gut-feel screening of a hundred resumes. The research does not hand you a clean verdict either way. Anyone who tells you the question is settled, in either direction, is ahead of the evidence. What is clear is that the result rides on the choices in this guide, not on the word "AI." The full evidence review is in Is AI hiring biased?
The short version
- AI has no fixed effect on bias. How you use it decides, and that part is yours.
- Stop asking "is AI biased?" Ask what job you are giving it, and on what information.
- Bias enters through known doors: training on past hires, scoring face and voice, black-box auto-rejection, automation bias, hidden proxies.
- Point the tool at what people can prove. Keep a human deciding. Refuse face and voice scoring. Demand a readable evidence trail.
- Know the law where you hire, and assume you are responsible for what your tools do.
- No tool is bias-free. Treat any vendor who claims otherwise as a warning sign.