Use case · Financial services
The same household, very different credit limits.
In 2019, applicants noticed that spouses with shared finances were being offered strikingly different Apple Card credit limits by gender, and the issuing bank, Goldman Sachs, faced a regulatory investigation. The lesson was not that someone typed "gender" into a formula. It is that the data carried gender anyway.
How bias hides in credit data
Remove the gender column and the signal remains: employment gaps correlate with parental leave; job roles, income patterns, and spending categories all carry it. These are proxies, features that stand in for the protected attribute. A model, or an analyst, reading them reconstructs gender without ever being shown it.
That is why "we do not use gender" is not a defence, and why regulators investigating discrimination look at outcomes, not column names.
What Rosa does
Rosa's Diagnose measures how recoverable gender is from the rest of the data, then Remove transforms the dataset so it cannot be recovered, while preserving each column's distribution. You name one protected attribute; Rosa finds the proxies itself.
This exact scenario is Test 1 in the Customer Portal: a credit dataset with proxy-encoded gender bias, debiased in the browser in one click, with the result framed the only honest way: no fairness/accuracy trade-off. The model trained on the debiased data performs as well, against the same unbiased ground truth, while the bias goes.
Who this is for
Credit risk, lending, and insurance pricing teams, and the compliance leaders accountable for them. Beyond the EU AI Act (credit scoring is named high-risk), GDPR anti-discrimination duties and gender-neutral pricing rules point at the same root cause: the data. Rosa fixes it there, and leaves the evidence.
Run the credit scenario yourself, today.
Test 1 is one click in the portal. Free. No credit card.