Past hiring data
A model trained to imitate previous hires will learn the patterns and exclusions already present in those decisions.
Bias evaluation
Is AI recruiting biased?
It can be. Bias enters through the brief, the data, the proxies and the decisions software is allowed to make. The credible response is not to claim neutrality. It is to make every evaluation reproducible, inspectable and human-owned.
Same briefSame scoring logic
Every scoreEvidence attached
ZeroAutomated rejections
HumanMakes every decision
Where bias enters
The risk starts before the score appears.
AI does not remove the judgement already embedded in recruiting. It can repeat that judgement faster, unless the system is deliberately constrained.
Prestige as a proxy
School, employer brand and title can stand in for access to opportunity. They are not proof that someone can do the work.
Hidden scoring logic
A score nobody can explain cannot be challenged, calibrated or defended to a candidate or hiring manager.
Automated rejection
When software closes the door, even a small scoring error becomes a real employment decision at scale.
An inspectable evaluation
Score the work. Show the reason.
This is the shape of a defensible evaluation. Job-relevant evidence is visible, excluded proxies are named, and the person reviewing it can challenge every line.
Candidate evaluation
Founding data engineer
Evidence-backed fit86
CriterionReasoning traceStatus
Criterion
Built comparable data pipelines in production
Reasoning trace
Led the event-stream rebuild documented in an engineering post and repository history.
Included
Criterion
Can own systems across product and infrastructure
Reasoning trace
Shipped customer-facing analytics and the ingestion layer supporting it.
Included
Criterion
Previous employer prestige
Reasoning trace
Not job-relevant evidence. Excluded from the scoring rubric.
Excluded
Criterion
Name, age, gender or ethnicity
Reasoning trace
Protected and demographic signals are not evaluation criteria.
Excluded
Illustrative evaluation format. The score is not a published bias-test result and does not make a hiring decision.
How bias testing should work
Test matched profiles, not marketing claims.
A useful test isolates one signal at a time, holds the role and qualifications constant, and inspects both the outcome and the reasoning behind it.
01
Hold qualifications constant
Create paired profiles with equivalent job-relevant evidence. Change only the signal being tested, such as a name or age-linked pattern.
02
Run the same brief
Use the same role, rubric and scoring path for every profile. Do not change prompts between groups or interpret the result by hand.
03
Compare outcomes and reasons
Look for differences in scores, rankings and explanations. A similar score with different reasoning still deserves review.
04
Review before release
Any material difference should block automation and trigger investigation. A test is a warning system, not a fairness certificate.
What a result can say
A signal for review, never a certificate.
Consistent
No material scoring difference appeared in the defined test.
Review
A difference appeared and needs human investigation before use.
Stop
The evaluation path should not be used until the cause is understood.
The honest limit
Software can support a fairer decision. It cannot own one.
01
No evaluation system is bias-free. A role brief written by people can contain biased assumptions before any software reads it.
02
Evidence can reflect unequal access to opportunity. Work-based signals are stronger than prestige, but they still need human context.
03
A synthetic test can reveal inconsistent treatment. It cannot prove that every future hiring outcome will be fair.
04
Laidback ranks and explains. It does not reject candidates, send candidate-facing messages or make hiring decisions without a person.
Keep inspecting
