AI and search
What explainable AI search should show a recruiter
Every recruitment product now has AI search. Very few of them can tell you why a candidate came third, which is the only thing that matters when a client asks.
General information. This is general information about how systems and processes work, written for Australian recruitment agencies. It is not legal, tax or financial advice. Where a decision turns on the Privacy Act, on licensing, on tax or business structure, on your contracts or on your obligations to candidates, take advice on your specific circumstances.
The problem with a score
A number between 0 and 100 next to a candidate’s name looks authoritative and cannot be argued with. That combination is the problem: it borrows the credibility of a measurement without any of the accountability of one.
A recruiter cannot defend it to a client. A candidate cannot challenge it. And nobody can tell whether it fell because the person lacks a skill or because a field was never filled in, which are opposite situations requiring opposite actions.
Show the interpretation before the results
A natural-language question has to become a structured query. That translation is where almost all the error is, and it is invisible in most products.
It should be shown as named filters with the words each one came from, before the search runs. "In Brisbane" becoming a 25km radius rather than all of South East Queensland is a decision the recruiter should see and be able to change in two seconds, not discover three shortlists later.
And the filters should be editable, because editing them is the search. A question you can only rephrase is a slot machine.
Four things a result must carry
Evidence: which field or document the match came from, linked, so it can be checked in one click.
Recency: when that evidence was recorded. A skill confirmed in 2019 is a different claim from one confirmed in June, and freshness is the most commonly ignored dimension in candidate data.
What is missing: named explicitly. "Work pattern not recorded, ask before shortlisting" is actionable. A silently lower rank is not.
Provenance: which source scope produced it, so a result from a widened scope is never mistaken for one of your own records.
What it must not do
It must not infer or use protected characteristics. Not to rank, not to filter, not to display. This is both a legal exposure and a quality problem: a model that has learned a proxy for age or gender from your historical placements will reproduce your historical bias with a confident number attached.
It must not contact anyone. Search finds people; a person decides to approach them, under the consent and preference controls on the record.
It must not search beyond the scope you set. "Everything we could find" is not a source scope, and a result whose origin cannot be named should not appear.
It must not decide. Ranking and explaining is useful. Selecting is a human act with a human accountable for it.
Four questions for a vendor
Show me the filters this question produced, before the results. If the answer is that it does not work that way, you cannot audit anything downstream.
Show me why this candidate ranked third, with the source of each factor.
What happens to a candidate missing a field the query asked about?
What data of mine is used to train models, and are those models shared with other customers? The only comfortable answer is that they are not.