A model that looks fair can quietly stop being fair. Here's how to catch it.

The question

A machine learning model is trained, tested, and its group-fairness numbers look fine. Then it's deployed. Does that fairness measurement survive contact with the real world — label noise, a mismeasured protected attribute, a subgroup nobody thought to check, a smaller-than-expected sample? Or does "fair on the test set" quietly stop meaning anything the moment reality gets messy?

What we actually did

Fairshift Lab is a synthetic laboratory — every number in it comes from a generator whose structural equation is written down and shown, not a black box. Version 1.3.0 added a Robustness Lab: two inspectable model families (logistic regression and a shallow decision tree), fixed before any results were generated, stress-tested under six controlled conditions — symmetric label noise, group-conditional label noise, protected-field measurement error, an unobserved intersectional subgroup, sample-size stress, and structural misspecification. Every hypothesis was written down in a preregistered protocol before a single result existed.

What we found

Nothing about this was a clean story, and that's the point:

• Both models degrade under stress, but not the same way on the same metric. A model family that holds up better on accuracy can degrade worse on fairness, and vice versa — there's no universal "more robust" model here.
• Group-conditional label noise — where one group's labels get flipped more often than another's — produced a real, measurable fairness gap that had nothing to do with the model's actual behavior toward that group. The unfairness was manufactured entirely by the label noise, not the model.
• A protected attribute measured with error made the observed fairness gap understate the true one — an audit using the recorded (noisy) attribute would have looked more reassuring than reality.

None of this says "AI fairness is fake" or "don't measure it." It says the opposite: a single fairness number, measured once, under clean conditions, is not evidence that fairness survives deployment. You have to stress-test it the same way you'd stress-test any other safety-critical measurement.

Try it yourself

The Robustness Lab lets you pick a stressor, drag its magnitude, and watch both models' accuracy, calibration, and fairness gaps move in real time — with the uncertainty and limitations of what you're looking at shown before any number, not after. If you want to go deeper: the full robustness report documents every hypothesis's actual disposition, including the ones that didn't confirm cleanly.

Learn more

• Hardt, Price & Srebro (2016), Equality of Opportunity in Supervised Learning, NeurIPS — the equal-opportunity/equalized-odds definitions this project measures against.
• Kleinberg, Mullainathan & Raghavan (2017), Inherent Trade-Offs in the Fair Determination of Risk Scores, ITCS — why you generally can't satisfy every fairness definition at once.
• The source code is MIT-licensed and fully reproducible — clone it, change the stressors, and see what breaks.