Disparate Impact: When One Group's Results Are Just Worse

Opening Scene

Two slides sit side by side under the microscope, prepared with the exact same procedure, drawn from the exact same protocol. One reads clean. The other shows a reading meaningfully worse than the first, for no reason the procedure itself can explain. Nobody set out to contaminate one slide more than the other — the process was, on paper, identical for both — but the outcome is unmistakably unequal, and that gap itself is the finding, regardless of anyone’s intent.

In Plain English

Disparate impact describes a situation where a facially neutral policy, process, or model produces substantially different outcomes for different groups, even without any explicit intent to discriminate and even without the protected attribute being used as a direct input. It’s typically measured using a rule of thumb like the “four-fifths rule”: if one group’s positive outcome rate falls below 80% of the group with the highest rate, that’s generally treated as evidence of disparate impact worth investigating further. The key insight is that intent doesn’t matter for disparate impact analysis — a model can be entirely “neutral” in its design and still produce a starkly unequal real-world result, and that result is what gets scrutinized.

The Old Way

Before disparate impact was treated as its own distinct, measurable category of harm:

Discrimination claims were often evaluated almost entirely on intent, making it very difficult to challenge a policy or model that was neutral on paper but unequal in practice.
A wide gap in outcomes between groups was sometimes dismissed as coincidental or explained away by unexamined “legitimate business factors” without rigorous testing of that explanation.
There was no widely agreed-upon threshold or rule of thumb for deciding when an outcome gap was large enough to warrant a genuine investigation.

Measuring the gap itself, independent of anyone’s intent, is what disparate impact analysis adds to a bias audit.

What’s Changing (and Why AI Is the Reason)

Disparate impact analysis has moved from a legal concept applied mostly to hiring practices into a standard, quantitative check applied broadly across AI-driven decisions.
This dovetails with the case-based reasoning covered in this content library’s dedicated data ethics case studies series, which documents real situations where a facially neutral system produced exactly this kind of unequal outcome.
As AI models increasingly make high-volume, high-stakes decisions at a scale where even a small per-decision gap compounds into a substantial aggregate disparity, disparate impact analysis has become an essential, standard part of any serious bias audit.

The Metaphor, Fully Extended

The Two Identical-Procedure Slides	Disparate Impact Concept
The same procedure applied to both samples	The same model or policy applied to both groups
No deliberate difference in how each was prepared	No explicit use of the protected attribute as an input
A meaningfully worse reading on one slide anyway	A meaningfully worse outcome rate for one group anyway
The gap itself being the finding, regardless of intent	The outcome gap itself being the finding, regardless of intent

For Beginners: What to Actually Do

Learn the basic idea of the four-fifths rule as a starting rule of thumb for spotting a concerning outcome gap.
Practice separating the question “was there bad intent” from the question “was there an unequal outcome” — they’re not the same question.
Get comfortable calculating simple outcome-rate ratios between groups as a first-pass check.

For Practitioners and Leaders: The Deeper Layer

Run disparate impact analysis as a standard part of every consequential model’s pre-deployment checklist, not just when a complaint prompts it.
Treat a flagged disparate impact ratio as the start of an investigation, not proof of a specific cause, since legitimate factors sometimes do explain part of a gap.
Document both the finding and the investigation into its cause, since regulators and courts increasingly expect to see that reasoning laid out.

Quick Recap

Disparate impact measures unequal outcomes between groups, regardless of intent or whether a protected attribute was a direct input.
The four-fifths rule is a common rule of thumb for flagging a gap worth investigating further.
A facially neutral model can still produce a starkly unequal real-world impact.
Disparate impact findings should trigger investigation into cause, not just documentation of the gap itself.

Where This Fits in the Series

Article 5 covered which attributes get tested for bias. This article covered what it looks like when one of those tested groups comes back with a meaningfully worse reading. Article 7 pulls all of these individual tests together into what a full, formal bias audit actually looks like end to end.