A recruiter rejecting one CV an hour on a hunch is a small, containable problem. You can coach that person. You can audit their decisions. You can sit them down and ask why every name they shortlisted sounds like their own. A model rejecting ten thousand CVs an hour on a pattern nobody can fully articulate is a different kind of problem entirely, and most of the companies buying these tools haven't worked out that it's worse.
That's the thing I keep saying in rooms where someone has just demoed a shiny "AI-powered talent platform" and the heads are nodding. Automating a process doesn't sanitise it. If the process was biased, you've now made the bias faster, cheaper, and far harder to see. You've taken something a human did clumsily and given it the gloss of mathematical objectivity, which is exactly the disguise discrimination wants.
Where the danger actually sits
I spent two decades inside large technology organisations before I moved into People work, and the engineering habit that stuck with me is this: be specific about which part of the system you're worried about. "AI in recruitment" is too broad to have an opinion on. So let me split it.
Ranking and scoring candidates is the dangerous end. A model trained on who your company hired and promoted in the past learns, with great fidelity, who your company hired and promoted in the past. If that history skewed male, or skewed towards one set of universities, or invisibly penalised a two-year gap that usually means childcare, the model absorbs all of it and presents the result as a score out of a hundred. The most cited example is still Amazon's experimental tool that taught itself to downgrade CVs containing the word "women's", but I don't lean on it because it's old — I lean on it because it's typical. The model did exactly what it was built to do. The training data was the discrimination.
Video "personality" scoring is worse, and I'll be blunt about it: I think most of it is closer to phrenology than science. Tools that infer conscientiousness or "culture fit" from facial micro-expressions, vocal tone, or word choice are making confident claims about interior states they cannot observe. They penalise accents, neurodivergence, a bad webcam, a candidate interviewing from a noisy flat because they share it with three other people. Illinois and a handful of other places now regulate exactly this for good reason. If you are scoring a human being's worth from how their face moves on a laptop camera, you have lost the plot.
Where it genuinely earns its place
I'm not against the technology. I'm against pointing it at the decision. Used on the work around the decision, it's straightforwardly useful.
- Scheduling. Coordinating five interviewers, two time zones and a candidate's day job is miserable admin, and there is no fairness question buried in a calendar invite. Automate all of it.
- Sourcing breadth. Search tools that surface candidates you'd never have found — different regions, non-obvious title histories, people who don't use the keywords your team happens to use — widen the top of the funnel. That's the opposite of bias, as long as the widening is where it stops and a person decides who to actually approach.
- Drafting and tidying. First-pass job descriptions, summarising a long CV so a human reads it faster, flagging that a posting is full of jargon that deters the very people you want. Assistive, reversible, low-stakes.
The pattern, if you want one line to carry out of here: let the machine widen the funnel and handle the logistics. Don't let it narrow the field. Breadth is safe. Judgement is not.
Somebody's name has to be on it
Here's the part people skip because it's inconvenient. For every consequential decision, a named human being has to be accountable — not the vendor, not "the algorithm", a person you could put in front of a tribunal. The moment a rejected candidate, or a regulator, or your own board asks "why was this person screened out", the answer cannot be a shrug at a black box. Under the EU AI Act, recruitment systems are high-risk and that accountability is becoming a legal requirement, not a nicety. But I'd want it even if no law demanded it, because a decision nobody will own is a decision nobody can defend.
So when I deploy any of this, the rules are dull and non-negotiable. The tool recommends; a person decides and signs. We keep records of why, in language a human wrote. We test outcomes across groups before go-live and on a schedule after — not because we expect a clean bill of health, but because the only honest assumption is that bias is present until measured otherwise. We tell candidates an automated step exists and give them a route to a human. And we keep one uncomfortable question pinned to the wall: if this model is wrong about someone, who finds out, and how?
Most vendors can't answer that last one. That's usually all I need to know.
The mess of it is that good hiring was always hard, slow, and human, and AI is being sold as the thing that finally makes it fast, cheap, and clean. It can make it faster and cheaper. Clean is the lie. If you remember nothing else: a tool that scales your hiring also scales whatever was already wrong with it, and "the system did it" has never once been an acceptable reason to end someone's chance at a job.
