What is adverse impact and the four-fifths rule?

Adverse impact occurs when a selection process passes one protected group at a substantially lower rate than another. The common four-fifths (80%) rule flags a problem when a group's selection rate is below 80% of the highest group's rate — a widely used screening test for disparate impact.

Why does the use stage matter?

A tool that ranks thousands of CVs has broad, automated reach, while a tool that suggests interview questions has a human firmly in the loop. The further a tool sits from final selection with human review, the lower its standalone fairness risk.

Should an AI tool be able to explain a rejection?

Yes, where it materially affects candidates. Many jurisdictions are moving toward a right to a meaningful explanation of automated decisions, and explainability is also your best defence if a hiring decision is challenged.

Does this tool save my answers?

No. The checker evaluates everything in your browser. Nothing you enter about the vendor or your answers is uploaded or stored.

What is the AI Recruitment Tool Fairness Checker?

Evaluate an AI recruitment tool against a structured fairness checklist — covering training data demographics, validation study design, adverse impact testing, explainability for rejected candidates, right to appeal, and bias monitoring — before you buy or deploy. It runs free in your browser on Gera Tools, with nothing uploaded.

AI Recruitment Tool Fairness Checker

Name: AI Recruitment Tool Fairness Checker
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

AI recruitment tool fairness checker

AI hiring tools can screen thousands of candidates in seconds — and replicate historical bias just as fast. Before you procure or deploy one, this checker walks you through the fairness questions that regulators, candidates, and your own legal team will eventually ask: where did the training data come from, was the tool validated, does it pass adverse-impact testing, and can a rejected candidate get an explanation and an appeal?

How it works

You describe the vendor tool and the stage it is used at — screening, interview scoring, or final selection — then answer a structured set of fairness questions grounded in the vendor’s documentation. The checker weights each answer by the tool’s use stage, because a tool that auto-rejects CVs carries far more standalone risk than one that merely suggests questions to a human interviewer. The result is a fairness risk rating plus a punch-list of the unanswered or failing items to put in front of the vendor before you sign.

Tips and notes

Demand the validation study. A reputable vendor can show that the tool predicts job performance and was tested for adverse impact; vague claims of “AI-powered matching” are a red flag.
Apply the four-fifths rule. Ask for selection rates by protected group; a group passing below 80% of the top group’s rate signals disparate impact that needs justification.
Insist on explainability and appeal for any tool that rejects candidates — it is both a fairness safeguard and your legal defence.
Re-test after every model update. Fairness is not a one-time procurement tick; the vendor changing the model can silently reintroduce bias.

The fairness questions that procurement often skips

Most procurement processes ask whether a tool works — does it predict job performance? — and skip the harder question: does it work equally well for all candidate groups? These are different questions, and a tool that scores strongly on predictive validity can still produce adverse impact if its training data over-represents historical hiring patterns that themselves were biased.

Training data demographics

The first question to put to any vendor: what was the demographic composition of the training data? If the tool learned from the resumes and outcomes of a workforce that was, say, 80% male in technical roles, its notion of a “good candidate” for a technical role will reflect that history. A vendor who cannot answer this question has not done the analysis. A vendor who has should be able to show you the composition and explain what steps were taken to address imbalances.

Validation design

Predictive validity studies measure whether a tool’s scores correlate with later job performance. The critical question is: validated on whom? A study conducted on current employees is not a study of rejected candidates — the people the tool would screen out are not in the dataset. This survivorship problem is endemic in AI hiring validation. Ask specifically whether the validation population was diverse across the protected groups relevant to your workforce.

The four-fifths rule in practice

The four-fifths (80%) rule works like this: if Group A passes a selection stage at a rate of 50% and Group B passes at a rate of 35%, Group B’s selection rate is 70% of Group A’s — below the 80% threshold — which flags potential disparate impact requiring examination. The rule is a screening test, not a legal verdict. Real scarcity (a role genuinely requiring a rare skill) can produce lower pass rates without discrimination; fabricated or proxy requirements cannot. When adverse impact is flagged, the question shifts to whether the requirement is job-related and consistent with business necessity.

Explainability and the right to appeal

Candidates rejected by an AI system have a legitimate interest in understanding why. Under UK GDPR and the EU AI Act’s high-risk provisions, meaningful explanation of automated decisions is increasingly a legal expectation rather than a nicety. “The algorithm ranked you below the threshold” is not a meaningful explanation. A tool that can identify the specific factors that contributed to a rejection — and explain them in terms a candidate can understand and challenge — is both fairer and legally more defensible.