AI in Recruiting — Frequently Asked Questions
Where AI helps in hiring and where it causes harm — automated screening, bias that redaction does not remove, the regulatory position, what to ask a vendor, and the false rejection rate nobody measures.
We have not built a recruiting system. See AI Recruiter for where we think the line is.
Screening#
Is AI screening not just faster and more consistent?#
Faster, certainly. "Consistent" is the word worth interrogating: a system applies the same judgement to every applicant, which is an improvement only if that judgement is sound.
There is an unintuitive consequence. Inconsistency in human screening acts as accidental protection — different reviewers have different blind spots, so a candidate rejected by one may be picked up by another. A single automated screen removes that variance along with everything else.
Why is screening the part you would build last?#
Because the failure is invisible. A rejected candidate never appears in your metrics, so a system quietly filtering out a category of applicant looks identical from the inside to one working well.
Everything else in hiring has a visible failure mode. This one does not.
Can bias be removed by hiding names and demographics?#
It helps and it is not sufficient. Proxies survive redaction — education, postcode, employment gaps, phrasing, non-native language patterns. Removing the obvious fields removes the obvious signals.
What about scoring candidates against our current high performers?#
It sounds rigorous and it reproduces the existing composition of your team by construction. If your current team is homogeneous for historical reasons, that is precisely what the model learns to select for.
Where it does help#
What is the strongest use?#
Structuring the interview: role-specific questions, a scoring rubric agreed in advance, a consistent format. Structured interviews are among the better-evidenced practices in hiring, and most organisations skip them because preparation is tedious.
That is the ideal automation target — real benefit, low harm, and it improves a human's input rather than replacing a human's decision.
What else is safe and useful?#
Writing better job descriptions, which measurably widens the applicant pool. Summarising applications into a consistent structure so a human compares like with like. Scheduling and candidate communication. Interview note-taking and evidence capture.
Note the boundary in the third: extracting and organising is different from scoring and ranking.
What about helping candidates?#
Practice questions and feedback are lower stakes, and the person affected is the one choosing to use it. See Interview Questions.
Rules and vendors#
What is the regulatory position?#
Hiring-related automated decision-making sits in the highest-risk categories of several regimes, with obligations around transparency, human oversight, record keeping and bias assessment. Some jurisdictions require independent bias auditing of automated employment tools and notice to candidates.
The practical point: compliance shapes the architecture and cannot be added afterwards. Confirm what applies to your jurisdictions with counsel — this is general information, not legal advice.
What should we ask a vendor?#
What does the system decide, and what does a human decide? Has it been independently audited for bias, and can we see the results? Was it trained on our past hiring decisions? Can a candidate learn an automated system was used and request review? What is the false rejection rate, and how was it measured?
Why is the false rejection rate the hard question?#
Because answering it requires assessing people you never hired, and no organisation has that data by default. You would need to advance a random sample the system rejected and see how they perform.
Almost nobody does this, so the honest answer is "unknown" — and a vendor willing to say that is more credible than one quoting a figure. If a number is offered, ask precisely how it was derived; the derivation is usually where the claim dissolves.
Practical#
What would you refuse to build?#
Automated rejection without a human decision. Ranking as the primary output — an ordered list invites the top ten to become the shortlist whatever the caveats. Video analysis inferring traits from expression or speech patterns. Scoring against a profile of current high performers.
Everyone is using AI to write applications now. What does that change?#
Applications become less informative and volume rises. The rational response is better structured evaluation rather than better filtering — which is the same conclusion as everything above.
Filtering harder on a noisier signal makes the noise worse.
Where would you start?#
A structured interview generator. Useful, improves a well-evidenced practice, produces an artefact humans use rather than a decision they rubber-stamp, and its worst failure is a mediocre question rather than someone not getting a job.