AI Hiring: Speed, Bias, and Who Gets Through the Door

Claire Monroe

Welcome to the show. Edwin, picture a recruiter opening a job post on Monday morning and finding 2,000 applications by lunch. The pitch for AI is obvious—just let the software handle the first pass. But the second I hear that, I think… okay, so now the software is deciding who even gets to exist in the process.

Edwin Carrington

Yes, and that’s the real shift. Not speed—control. Once AI is used to source, screen, rank, or even make decisions, it’s no longer just handling admin work. It’s shaping the pathway itself. The 2025 ILO paper on AI in HR makes this very clear—these systems rely on three things: objectives, data, and programming. And each of those can quietly drift away from what actually makes a good hire.

Claire Monroe

Those three—objectives, data, programming—they sound neat. Almost suspiciously neat. So if something breaks, where does it usually go first? Is it the goal, the data, or the way it’s built?

Edwin Carrington

Usually the objective. Because the objective has to be translated into something measurable. An employer says, “I want potential” or “good judgment,” but a system can’t optimize for abstract qualities. It needs a proxy. And that’s where things start to slip. If the target is vague, the data incomplete, or the model simplistic, the system can still be perfectly consistent… and consistently wrong.

Claire Monroe

So “good candidate” slowly turns into… whatever the system can count. And then it starts to feel less like you’re discovering talent and more like you’re defining it after the fact.

Edwin Carrington

Exactly. There’s a concrete example from a Canadian government pilot. They tried to assess “growth mindset” using recorded responses, and one approach reduced that concept to keyword frequency. Not context, not reasoning—just how often certain words appeared.

Claire Monroe

Keyword frequency… for growth mindset. That’s—yeah, that’s almost impressive in the worst way. Like measuring leadership by how many times someone says “alignment.”

Edwin Carrington

It shows the gap between intention and execution. The language at the top stays aspirational, but underneath, the system needs something rigid. And as application volume increases—which it has—organizations automate more aggressively. That creates a paradox. More applicants push efficiency, and efficiency increases distance from the actual person.

Claire Monroe

And that volume isn’t random. It’s platform-driven. One-click applications, easy apply—people send out dozens, sometimes hundreds. So companies see the flood and think, “we need automation,” but the automation is reacting to a problem the system itself helped create.

Edwin Carrington

Exactly. And faster systems are not necessarily better systems. The ILO paper warns against assuming objectivity just because something is digitized. A model may be consistent, but consistent against what? If it’s screening for narrow patterns—certain schools, continuous employment, specific phrasing—it’s just formalizing a guess about talent.

Claire Monroe

So the real risk isn’t just mistakes. It’s scaling a half-formed definition of merit across thousands of people without ever questioning it.

Edwin Carrington

That’s exactly it. And this is where the OECD’s work on trustworthy AI becomes relevant. Their position isn’t anti-technology. It’s about accountability. Systems that influence people’s opportunities should be transparent, understandable, and open to challenge. If a tool affects someone’s chances, you should be able to explain how and why.

Claire Monroe

“Open to challenge” is the part that sticks. Because if a human rejects me, I can at least imagine a flawed judgment. But if a system filters me out and no one can explain it… that’s not just frustrating. It’s opaque.

Edwin Carrington

And that lack of transparency is what erodes trust very quickly.

Claire Monroe

So let’s not turn this into a lazy “AI is bad” argument. Where does it actually help? Where does it make sense?

Edwin Carrington

The most effective uses are narrow and supportive. Summarizing applications so recruiters can process volume faster. Organizing pipelines. Identifying patterns in large datasets. Supporting compensation benchmarking. These reduce workload without pretending to assess deeper qualities like judgment or potential.

Claire Monroe

So it works as an assistant. Not as the final decision-maker.

Edwin Carrington

Exactly. Problems emerge when systems start making or heavily influencing high-stakes decisions with weak evidence. The ILO paper highlights risks like proxy discrimination, overly simplistic filters, and tools that claim to interpret facial expressions or tone of voice without proving relevance to job performance.

Claire Monroe

And if a tool claims it can read micro-expressions and predict success… the correct reaction is not curiosity. It’s skepticism.

Edwin Carrington

Yes. The burden of proof matters. If a system claims predictive power, it should demonstrate it clearly and specifically for the role in question.

Claire Monroe

Otherwise we’re just dressing up old pseudoscience with better UI.

Edwin Carrington

That’s not an unreasonable comparison. And then there’s ad delivery, which sounds neutral but often isn’t. Research using Meta’s Ad Library has shown that job ads can be distributed unevenly across demographics, even without explicit targeting.

Claire Monroe

Which means bias can happen before anyone even applies. If certain groups don’t see the opportunity in the first place, the funnel is already skewed.

Edwin Carrington

Exactly. And that’s an important point. Trustworthy AI isn’t just about the final decision. It’s about the entire system—who sees the opportunity, who gets screened, how decisions are made, and whether those decisions can be explained.

Claire Monroe

And then companies look at the final pool and say, “well, that’s just who applied.” Without realizing the system helped shape that pool.

Edwin Carrington

Yes. Which is why the key question isn’t “can AI do this?” It’s “what decision is it making, based on what data, and can that decision be challenged?” If those answers aren’t clear, then the organization isn’t controlling the system—the system is controlling the organization.

Claire Monroe

That’s… not a comfortable thought. Because buying software feels like progress, but it can actually be outsourcing judgment.

Edwin Carrington

It is a transfer of judgment. And managing that requires discipline. Keep humans involved in meaningful decisions. Test for unintended impact regularly. Be explicit about what the system is optimizing for. Involve people who understand the role, not just the technology.

Claire Monroe

And if someone gets filtered out, there should at least be a way to question it. Because “computer says no” isn’t a hiring strategy.

Edwin Carrington

No, it isn’t. The real measure of success isn’t just speed or efficiency. It’s whether hiring improves without narrowing opportunity. If the process becomes cleaner but the pool becomes smaller or less diverse, then the system has optimized the wrong outcome.

Claire Monroe

Faster hiring is easy to track. Reduced opportunity… not so much. That’s where things quietly go wrong.

Edwin Carrington

And the organizations that recognize that early tend to make better decisions—not just faster ones.