Can AI screen candidates fairly? What small businesses need to know

A study published by the Brookings Institution tested three large language models against 554 resumes carrying 80 different names signaling race and gender, matched against 571 real job descriptions. White-associated names were preferred in 85.1% of the head-to-head comparisons. Black-associated names won in 8.6%. The two came out roughly equal in just 6.3% of tests.

That’s the honest starting point if you’re asking whether AI can screen candidates fairly. Not “yes, it’s neutral.” Not “no, never trust it.” That number is what happens when a screening tool ranks people and nobody checks how the ranking got made.

You didn’t sign up to become an expert in algorithmic fairness. You signed up to run a business, and hiring landed on your desk anyway. You’re the one about to trust a piece of software with 40 resumes for a role you need filled by Friday, and you deserve a straight answer instead of a sales page.

Here’s ours. Whether AI screening is fair depends less on the algorithm than on what you can see when you review it: why a candidate scored the way they did, not just the score itself. A reviewer who only sees a number defers to a biased tool almost as often as if no reviewer were there at all.

No screening algorithm starts out neutral

Where the bias comes from

Every AI screening tool learns its sense of a “good” candidate from somewhere. Usually that’s a mix of the job description, the criteria an employer feeds it, and patterns in resumes and hiring outcomes it was trained or tuned on. None of that data is neutral. It reflects who got hired before, which reflects whatever bias already existed in that hiring.

Researchers at Princeton and the University of Chicago found this pattern gets worse, not better, when you hand a language model the hiring decision itself. In a simulated hiring game reported by MIT Technology Review, models including GPT, Claude, and Gemini variants sorted fictional candidates into job categories by ethnicity after just a handful of early observations, even though every group had identical success rates. On a segregation scale where human participants in the same test scored 0.84, the models scored roughly 65% higher. One model hit 1.83, close to the maximum possible score.

The industry already knows this

None of that means you should swear off AI screening. It does mean you should be suspicious of anyone who tells you their AI is simply “unbiased” or “objective.” A Resume Builder survey of 948 business leaders found that 83% of companies expected to use AI to screen resumes, and 67% of those same companies admitted their own tools could introduce bias. The industry mostly knows this. It just doesn’t always say it out loud when it’s selling you something.

If you want the fuller picture of where AI genuinely helps in hiring and where it doesn’t, we’ve written more on AI in hiring generally.

Why “just add a human reviewer” doesn’t fix it

What the deference study found

The standard advice, from compliance guides to law firm blog posts, is to keep a human in the loop. It’s true advice. It’s also incomplete in a way that matters a lot for how you actually use a tool.

A University of Washington study had 528 people make hiring decisions with AI systems showing different levels of racial bias, across resumes for 16 real job categories. When there was no AI involved, people picked candidates from different racial groups at close to equal rates. When the AI showed severe, measurable bias, people followed its recommendation about 90% of the time anyway.

A rubber stamp isn’t oversight

The researchers’ explanation was simple: unless the bias was obvious, people accepted the AI’s judgment. And a tool that hands you a match score or a “recommended” flag, with no visible reasoning attached, makes bias very hard to spot. You’re not reviewing the candidate. You’re reviewing the AI’s opinion of the candidate, and rubber-stamping it.

That’s not a reason to distrust every tool that ranks candidates. It’s the reason “human in the loop” only works when the human can see what the loop is actually judging. Anonymizing resumes doesn’t solve this on its own either, which we’ve covered in more depth in does blind hiring still make sense.

Fairness lives in what you can see, not in the algorithm

Here’s the reframe worth sitting with. You cannot certify an algorithm as fair the way you’d certify a scale as accurate. What you can build is a workflow where bias has somewhere to get caught.

That means three things have to be true at once. Your own criteria drive the score, not a hidden formula the vendor won’t explain. Every candidate gets measured against those same criteria, consistently, the way most candidate screening software claims to work. And you can see the reasoning behind any score, not just the number, so you can catch a ranking that doesn’t hold up and override it.

This is the actual mechanism behind the “AI surfaces, you decide” idea, not just a nice line. Surfacing the evidence is what makes the human decision still a human decision, instead of a rubber stamp on a score you never really evaluated. For the fuller checklist we use when vetting a tool against this standard, see responsible AI in hiring.

A consistent process still isn’t automatically a fair one

You might reasonably think: my process doesn’t need to be perfectly bias-free, it just needs to be more consistent than me skimming resumes at 11pm after a full day running the place. That’s a fair instinct, and it’s not wrong. It’s just not the whole answer.

Consistency and fairness aren’t the same thing. A process that scores every resume the same biased way is still unfair. It’s just uniformly unfair, which can look tidier on paper and be harder to notice, because nobody’s score jumps around for no reason. Resume screening software built well gives you consistency. Consistency without visibility just hides bias evenly across everyone instead of unevenly.

The fix is making sure the consistent thing you’re looking at is your own criteria, applied the same way every time, with the “why” attached so you can catch it when the criteria produce a bad call.

What this looks like in a real 40-resume pile

Setting your own criteria first

Say you’re hiring a front-desk coordinator. This is how we’ve built Truffle, our own candidate screening platform, to handle exactly this: resume screening, one-way video interviews, and talent assessments, combined however a role needs. You post the role Monday morning and by Wednesday you have 40 resumes, a handful of completed one-way video interviews, and maybe a couple of assessment results if you’re using them for this hire.

Before any of that happens, you tell the system what actually matters for this role. The must-haves, the deal-breakers, what good communication looks like for a front desk that fields walk-ins all day. That’s your standard, not a formula baked in from someone else’s hiring history.

What you actually see when you review

AI reads the pile against that standard. It scores each candidate’s match percentage, and next to that number it shows you why: which resume lines it pulled the experience from, which part of a recorded answer it flagged for communication, where an assessment result lines up or doesn’t with what you said you needed. You get a ranked list and the evidence behind every spot on it, together.

If a candidate ranks lower than you expected, you can open their profile, read the actual answer, and decide the AI missed something. You can move them up. Nobody’s stopping you, and nothing forces you to take the AI’s word for it. It doesn’t promise a bias-free outcome. No honest tool can. It gives you the visibility to catch a bad call and the standing to overrule it, which is the part most tools skip. This is what recruitment software built for a small business should actually look like: a score you can open up and check, not one you’re asked to just trust.

Can AI screen candidates fairly? Ask this before you buy

So, can AI screen candidates fairly? The honest answer is that no AI can promise fairness on its own, and you should be skeptical of anyone who claims theirs does. What it can do is apply the criteria you set, the same way, to every single candidate, and show its work clearly enough that you’re the one making the call, not the algorithm.

That reframes the question you should actually be asking a vendor. Not “is your AI unbiased,” because nobody can prove a negative like that. Ask instead: can you show me why, every time, for every candidate? And when I disagree with the score, is there anything stopping me from acting on that? If a tool can’t answer both of those clearly, the score it hands you is just a percentage with nothing behind it, regardless of what the sales page says about fairness. It’s a fair question to ask at any pricing tier, from a $49/month starter plan up.

If you’re hiring in a state or country with specific AI hiring rules already on the books, that changes what you need to document, not what you need to believe. Worth a look: our breakdown of the EU AI Act and hiring if any of your hiring touches EU candidates.

Frequently asked questions about AI and fair candidate screening

Can AI screening eliminate bias in hiring?

No. No AI system can honestly claim to eliminate bias, and you should treat any vendor who says so as a red flag. What a well-built tool can do is apply the same, employer-defined criteria consistently to every candidate and show the reasoning behind each score, so a human reviewer has something real to check instead of just a number.

Is AI resume screening regulated for small businesses?

Increasingly, yes, though the rules vary a lot by location. Some U.S. cities and states, and the EU under the AI Act, now require disclosure, bias auditing, or human oversight for automated hiring tools. If you hire across state lines or internationally, it’s worth checking what applies to you specifically rather than assuming a small team is exempt.

Does the fairness problem apply to one-way video interviews too, or just resumes?

The same principle applies. AI can transcribe and analyze a recorded answer and flag what stood out, but it shouldn’t be presenting a verdict on the candidate. Look for tools that show you the actual clip or transcript behind any summary, so you’re evaluating the candidate’s answer, not an AI’s opinion of it.

What should I actually ask a vendor to check if their AI screening is fair?

Ask two things. Can you see the reasoning behind every score, not just the score itself? And is there anything in the product that stops you from overriding a ranking you disagree with? If the answer to either is unclear, that’s the real signal, more than anything printed on the pricing page.