Latest / The Science of Leading / AI Hiring: Speed, Bias, and Who Gets Through the Door

Transcript

AI Hiring: Speed, Bias, and Who Gets Through the Door · The Science of Leading

• 0:00
  Welcome to the show.
• 0:01
  Edwin, picture a recruiter opening a job post on Monday morning and finding
• 0:05
  2, 000 applications by lunch.
• 0:07
  The pitch for AI is obvious-just let the software handle the first pass.
• 0:11
  But the second I hear that, I think... okay,
• 0:14
  so now the software is deciding who even gets to exist in the process.
• 0:14
  Yes,
• 0:20
  and that's the real shift.
• 0:22
  Not speed-control.
• 0:24
  Once AI is used to source, screen,
• 0:27
  rank, or even make decisions, it's no longer just handling admin work.
• 0:31
  It's shaping the pathway itself.
• 0:33
  The 2025 ILO paper on AI in HR makes this very clear-these systems rely
• 0:40
  on three things: objectives, data,
• 0:43
  and programming.
• 0:44
  And each of those can quietly drift away from what actually makes a good
• 0:49
  hire.
• 0:49
  Those three-objectives, data,
• 0:51
  programming-they sound neat.
• 0:53
  Almost suspiciously neat.
• 0:55
  So if something breaks, where does it usually go first?
• 0:57
  Is it the goal, the data, or the way it's built?
• 0:57
  Usually the objective.
• 1:02
  Because the objective has to be translated into something measurable.
• 1:06
  An employer says, "I want potential" or "good judgment," but a system
• 1:10
  can't optimize for abstract qualities.
• 1:13
  It needs a proxy.
• 1:14
  And that's where things start to slip.
• 1:16
  If the target is vague, the data incomplete,
• 1:19
  or the model simplistic, the system can still be perfectly consistent...
• 1:23
  and consistently wrong.
• 1:23
  So "good candidate" slowly turns into... whatever
• 1:29
  the system can count.
• 1:30
  And then it starts to feel less like you're discovering talent and more
• 1:33
  like you're defining it after the fact.
• 1:33
  Exactly.
• 1:36
  There's a concrete example from a Canadian government pilot.
• 1:40
  They tried to assess "growth mindset" using recorded responses,
• 1:44
  and one approach reduced that concept to keyword frequency.
• 1:47
  Not context, not reasoning-just how often certain words appeared.
• 1:47
  Keyword
• 1:53
  frequency... for growth mindset.
• 1:55
  That's-yeah, that's almost impressive in the worst way.
• 1:58
  Like measuring leadership by how many times someone says "alignment."It
• 2:03
  shows the gap between intention and execution.
• 2:06
  The language at the top stays aspirational,
• 2:08
  but underneath, the system needs something rigid.
• 2:11
  And as application volume increases-which it has-organizations automate
• 2:17
  more aggressively.
• 2:18
  That creates a paradox.
• 2:20
  More applicants push efficiency,
• 2:22
  and efficiency increases distance from the actual person.
• 2:22
  And that volume
• 2:27
  isn't random.
• 2:28
  It's platform-driven.
• 2:29
  One-click applications, easy apply-people send out dozens,
• 2:33
  sometimes hundreds.
• 2:34
  So companies see the flood and think,
• 2:36
  "we need automation," but the automation is reacting to a problem the
• 2:40
  system itself helped create.
• 2:40
  Exactly.
• 2:43
  And faster systems are not necessarily better systems.
• 2:47
  The ILO paper warns against assuming objectivity just because something
• 2:52
  is digitized.
• 2:53
  A model may be consistent, but consistent against what?
• 2:57
  If it's screening for narrow patterns-certain schools,
• 3:01
  continuous employment, specific phrasing-it's just formalizing a guess
• 3:05
  about talent.
• 3:05
  So the real risk isn't just mistakes.
• 3:09
  It's scaling a half-formed definition of merit across thousands of people
• 3:12
  without ever questioning it.
• 3:12
  That's exactly it.
• 3:16
  And this is where the OECD's work on trustworthy AI becomes relevant.
• 3:21
  Their position isn't anti-technology.
• 3:24
  It's about accountability.
• 3:25
  Systems that influence people's opportunities should be transparent,
• 3:29
  understandable, and open to challenge.
• 3:31
  If a tool affects someone's chances,
• 3:33
  you should be able to explain how and why."Open to challenge" is the
• 3:38
  part that sticks.
• 3:39
  Because if a human rejects me,
• 3:41
  I can at least imagine a flawed judgment.
• 3:43
  But if a system filters me out and no one can explain it... that's not
• 3:46
  just frustrating.
• 3:47
  It's opaque.
• 3:47
  And that lack of transparency is what erodes trust very quickly.
• 3:47
  So
• 3:53
  let's not turn this into a lazy "AI is bad" argument.
• 3:56
  Where does it actually help?
• 3:57
  Where does it make sense?
• 3:57
  The most effective uses are narrow and supportive.
• 4:02
  Summarizing applications so recruiters can process volume faster.
• 4:07
  Organizing pipelines.
• 4:09
  Identifying patterns in large datasets.
• 4:12
  Supporting compensation benchmarking.
• 4:14
  These reduce workload without pretending to assess deeper qualities like
• 4:19
  judgment or potential.
• 4:19
  So it works as an assistant.
• 4:22
  Not as the final decision-maker.
• 4:22
  Exactly.
• 4:25
  Problems emerge when systems start making or heavily influencing high-stakes
• 4:31
  decisions with weak evidence.
• 4:32
  The ILO paper highlights risks like proxy discrimination,
• 4:37
  overly simplistic filters, and tools that claim to interpret facial expressions
• 4:42
  or tone of voice without proving relevance to job performance.
• 4:42
  And if a
• 4:46
  tool claims it can read micro-expressions and predict success... the correct
• 4:50
  reaction is not curiosity.
• 4:51
  It's skepticism.
• 4:51
  Yes.
• 4:54
  The burden of proof matters.
• 4:56
  If a system claims predictive power,
• 4:58
  it should demonstrate it clearly and specifically for the role in question.
• 4:58
  Otherwise
• 5:03
  we're just dressing up old pseudoscience with better UI.
• 5:03
  That's not an
• 5:07
  unreasonable comparison.
• 5:09
  And then there's ad delivery, which sounds neutral but often isn't.
• 5:13
  Research using Meta's Ad Library has shown that job ads can be distributed
• 5:18
  unevenly across demographics, even without explicit targeting.
• 5:18
  Which means
• 5:23
  bias can happen before anyone even applies.
• 5:26
  If certain groups don't see the opportunity in the first place,
• 5:29
  the funnel is already skewed.
• 5:29
  Exactly.
• 5:31
  And that's an important point.
• 5:33
  Trustworthy AI isn't just about the final decision.
• 5:37
  It's about the entire system-who sees the opportunity,
• 5:41
  who gets screened, how decisions are made,
• 5:43
  and whether those decisions can be explained.
• 5:43
  And then companies look
• 5:48
  at the final pool and say, "well,
• 5:49
  that's just who applied." Without realizing the system helped shape that
• 5:53
  pool.
• 5:53
  Yes.
• 5:54
  Which is why the key question isn't "can AI do this?" It's "what decision
• 5:59
  is it making, based on what data,
• 6:01
  and can that decision be challenged?" If those answers aren't clear,
• 6:06
  then the organization isn't controlling the system-the system is controlling
• 6:10
  the organization.
• 6:10
  That's... not a comfortable thought.
• 6:13
  Because buying software feels like progress,
• 6:16
  but it can actually be outsourcing judgment.
• 6:16
  It is a transfer of judgment.
• 6:20
  And managing that requires discipline.
• 6:23
  Keep humans involved in meaningful decisions.
• 6:26
  Test for unintended impact regularly.
• 6:29
  Be explicit about what the system is optimizing for.
• 6:33
  Involve people who understand the role,
• 6:36
  not just the technology.
• 6:36
  And if someone gets filtered out,
• 6:40
  there should at least be a way to question it.
• 6:42
  Because "computer says no" isn't a hiring strategy.
• 6:42
  No,
• 6:47
  it isn't.
• 6:47
  The real measure of success isn't just speed or efficiency.
• 6:51
  It's whether hiring improves without narrowing opportunity.
• 6:55
  If the process becomes cleaner but the pool becomes smaller or less diverse,
• 7:00
  then the system has optimized the wrong outcome.
• 7:00
  Faster hiring is easy
• 7:05
  to track.
• 7:05
  Reduced opportunity... not so much.
• 7:08
  That's where things quietly go wrong.
• 7:08
  And the organizations that recognize
• 7:12
  that early tend to make better decisions-not just faster ones.