A bias check is not a philosophical exercise. It is arithmetic on your own hiring data, and you can do a useful version in an afternoon.
What it measures
Not what anyone intended. What actually happened.
Specifically: does the proportion of people who get through a stage differ between groups, and by how much?
Two numbers do the work:
- Pass rate — what share of a group got through a stage.
- Ratio — one group's pass rate divided by the pass rate of the group that did best.
In US practice a common rule of thumb treats a ratio below four-fifths as worth looking into. It is a starting point for asking questions, not a legal line, and not proof of anything either way.
How to run it
- Pick a stage with a clear pass or fail. Screening is the obvious one.
- For each group, count how many entered and how many got through. Divide one by the other.
- Find the highest pass rate. Divide every other group's rate by it.
- Write the actual counts next to the ratios. Always.
- Repeat for each stage, then for the whole process from start to finish.
The whole-process version matters, because small differences at four stages add up to a large one overall while no single stage looks alarming.
Reading the result honestly
| What you find | What it does and does not mean |
|---|---|
| Low ratio, large numbers | A real difference. Find the rule causing it |
| Low ratio, tiny numbers | Might be chance. Report the counts. Do not declare innocence or guilt |
| All ratios near 1 | No difference at this stage. Not proof the process is fair — check earlier stages |
| Difference only shows up overall | Small gaps adding up. The most commonly missed pattern |
Two traps. A clean result at interview stage tells you nothing if the difference happened at the search stage, because the affected people were never in the pool. And in most companies the group sizes are small enough that a handful of people moves a ratio — which is a reason to show the counts and combine several months, not a reason to dismiss what you found.
Check how well the CVs were read, too
Almost every bias check skips this, and Module 4 explained why it matters.
Compare how well your tool reads CVs across groups and across file formats. If it fails more often on one kind of CV template, one alphabet, or one style of career history, that is people being removed by a technical fault before any scoring happens.
It will never show up in a check that only looks at scores, because the affected people simply have worse data.
In practice: take the CVs that were read badly, look at what they have in common, and ask whether that thing is spread evenly across your applicants. It usually is not.
Who should do it
Where the law requires an independent check — New York City is the clearest example — independence has a specific meaning and your own review will not satisfy it.
Everywhere else, do it yourself, regularly. Treat any formal audit as a separate obligation rather than a replacement. A yearly external audit tells you about last year. A quarterly internal one tells you before it becomes a pattern.
Write it down
Record the date, the period covered, the stage, the groups and counts, how you did it, what you found, what you concluded and what you changed.
A check with no recorded action is evidence that you knew — which is worse than not having looked.
Download the risk register · all templates
What to do when you find something
Look at the rule, not the people. A difference usually traces back to one requirement acting as a stand-in for something else — years of experience, unbroken employment, which university, location. Change the rule and run it again.
Adjusting individual outcomes to fix the number treats the symptom and creates a new problem.