Skip to content
All apps

Bias in candidate screening

About this app

A model trained on past hiring decisions sorts 200 applications. Delete the group from the data and the ratio barely improves, because a proxy feature carries the bias on.

With explain mode: an AI tutor walks you through the app, points out what matters in the scene and answers your questions.

The problem

When a model learns from past hiring decisions, it inherits their biases. If one group was disadvantaged in the past, the model carries that disadvantage forward. Simply removing the protected attribute is often not enough: other features that correlate with the group act as stand-ins, or proxies. One common measure is the four-fifths rule: if one group’s selection rate is below 80 % of the other group’s rate, that is a warning sign. A twin test adds a second check, asking whether equally qualified people receive the same decision.

What the app shows

200 fictional profiles, 100 in group A and 100 in group B, with exactly the same distribution of qualification and experience. A small model, a logistic regression, is trained in the browser on biased past decisions and hands out 60 invitations. Three steps lead from “Historical” to “Without group” to “Without proxy”. The feature “gap in CV” is much more common in group B and takes over the role of the group. Metrics show invitation rates, the four-fifths ratio, the twin test and a history. A guided explanation mode walks through the scene.

Try it yourself

First switch off the group feature and notice that the rates barely improve. Only without the CV gap do the groups move closer. Then compare two interventions: relabelled training data that rates only qualification and experience, or a separate threshold per group with 30 invitations each. Tap a figure to see its twin with the same qualification, and check whether both receive the same decision.

Frequently asked questions

What is AI bias in recruiting?
Bias here means that a model systematically disadvantages applicants based on a feature unrelated to their suitability. A common cause is training data from earlier decisions that were already biased. The model learns these patterns and applies them to new applications.
What is a proxy variable?
A proxy is a feature that looks harmless on its own but is strongly linked to a protected attribute, for example a gap in the CV that occurs more often in one group. If the protected attribute is removed, the model can carry the same bias through the proxy. That is why “fairness through unawareness” usually falls short.
What is the four-fifths rule?
The four-fifths rule compares the selection rates of two groups. You divide the lower rate by the higher one; a ratio below 0.8 is treated as an indication of adverse impact. It is a rule of thumb for detection and does not replace a case-by-case assessment.
What can I do in the bias screening app?
You switch features such as group or CV gap on and off, rerun the selection and compare two countermeasures. Metrics show invitation rates, the four-fifths ratio and the twin test. All profiles are fictional, and the model is trained in your browser.

Subject: AI literacy

More from AI literacy

A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps