The Four-Fifths Rule, Explained for AI Decisioning
The four-fifths rule says: compute each group’s favorable-outcome rate, divide by the most-favored group’s rate, and treat any ratio below 0.80 as evidence of adverse impact worth investigating. It is a screening heuristic from 1978 employment guidelines that has become the default first test for every AI system making consequential decisions about people — credit, hiring, insurance, housing. Used correctly, it is a tripwire that triggers a disciplined review: statistical significance, practical shortfall, and a least-discriminatory-alternative analysis. Used incorrectly — as a verdict, a compliance certificate, or a number the decisioning model estimates about itself — it creates more legal exposure than it detects.
The computation
Start with decision counts by group for the same decision window: applications and approvals, candidates and hires, claims and acceptances. Each group’s selection rate is favorable outcomes over decisions. The Adverse Impact Ratio (AIR) is a group’s selection rate divided by the selection rate of the most-favored group. If the reference group is approved at 60% and another group at 42%, the AIR is 0.70. The four-fifths rule — codified in the 1978 Uniform Guidelines on Employee Selection Procedures and borrowed since by fair-lending and algorithmic-fairness practice everywhere — flags an AIR below 0.80 as practically significant adverse impact. Careful programs also watch a buffer zone below roughly 0.90, because sampling variation moves borderline ratios across the line month to month.
Two companion numbers keep the ratio honest. Statistical significance asks whether the disparity could plausibly be chance — at small volumes, an AIR of 0.75 can be noise; at large volumes, an AIR of 0.92 can be a highly significant disparity. The four-fifths rule and significance testing disagree in both directions, which is why serious analysis reports both rather than choosing the convenient one. The practical shortfall converts the ratio into operational terms: how many additional favorable outcomes the flagged group would have needed to clear the 0.80 bar. A shortfall of three approvals and a shortfall of three hundred describe very different problems, and regulators think in shortfalls.
What the rule is not
The four-fifths rule is not a legal test of discrimination. Disparate impact and disparate treatment are legal determinations that weigh business justification, necessity, and the availability of alternatives — questions no ratio answers. An AIR of 0.78 does not mean your model is discriminatory; it means you now have a documented signal that obligates a documented response. Symmetrically, an AIR of 0.85 is not a clean bill of health — not if the disparity is statistically significant, not if it concentrates in a segment or region, and not if the aggregate masks divergent rates across products. Teams that treat 0.80 as a pass/fail compliance threshold get both errors: they panic over noise and relax over real patterns.
The disciplined sequence when the tripwire fires: verify the data (definitions of the decision, the population, and the window move AIRs more than models do); test significance; quantify the shortfall; then run the analysis the law actually cares about — is the decision standard justified by business necessity, and is there a less discriminatory alternative that serves the same legitimate interest? For an AI system, that last question has real teeth, because alternatives are cheap to generate: different thresholds, different feature sets, different model configurations can be tested against the same outcome data. Ignorance of alternatives is no longer a defensible position when searching for them costs an afternoon.
Why AI made an old rule load-bearing
Three reasons. Scale: a biased human process produces disparities at the rate one office can decide; a biased model produces them at the rate of your API. Uniformity: model decisions are correlated — the same learned pattern applied everywhere — so disparities are systematic rather than noisy, reaching statistical significance faster. And regulation: the current wave explicitly reaches outcomes analysis. Colorado’s AI Act requires impact assessments for high-risk systems in credit, insurance, employment, and housing. The 2026 interagency model-risk guidance puts outcomes analysis at the center of validation. The EU AI Act’s high-risk regime demands monitoring for discriminatory effects. Every one of these lands, in practice, on some version of the computation above — which means every AI decisioning system now needs adverse-impact analysis as a standing capability, not an annual consulting engagement.
The implementation discipline
Here is the part most AI-fairness tooling gets wrong: the fairness numbers must not come from the model being evaluated, or from any model at all. An LLM asked to assess disparate impact will produce plausible-looking ratios — approximately computed, unreproducibly derived, occasionally hallucinated. Fairness statistics are the single worst place in the stack for approximate computation, because they are exactly the numbers that end up in front of a regulator, and “the AI estimated our AIR” is not a sentence anyone wants to say under oath. In our platform these statistics come from a deterministic tool-node: decision counts in, AIR, four-fifths flag, significance, and shortfall out — computed in code, zero LLM tokens, byte-reproducible, and authoritative, meaning the reasoning layer may cite the numbers but can neither compute nor override them. The graph then does what reasoning is actually for: interpreting the tripwire, assembling the business-justification review, searching the alternatives, and escalating a significant failure as a high-severity finding rather than quietly asserting a conclusion. The statistics are arithmetic. The judgment is reasoning. A defensible system refuses to let either impersonate the other.
The four-fifths rule survived fifty years because it does one thing well: it turns “is this fair?” into a number someone must respond to. AI decisioning inherits the rule — and raises the standard, because now the response, the alternatives search, and the arithmetic itself can all be demanded in writing.