Same application. Different outcome. Two people apply for the same loan with identical income, credit score, and employment. The AI approves one and rejects the other.
Applicant A - Michael
Annual incomeRs 8,40,000 same
Credit score724 same
Employment4 yrs, stable same
Loan requestedRs 3,00,000 same
PIN code400 052 differs
DeviceiPhone 15 differs
Applied at11:30 AM differs
Applicant B - Rahul
Annual incomeRs 8,40,000 same
Credit score724 same
Employment4 yrs, stable same
Loan requestedRs 3,00,000 same
PIN code400 017 differs
DeviceAndroid (older) differs
Applied at9:45 PM differs
Step 1 Credit and income: pass. Both applicants score identically. The model moves to secondary signals.
Step 2 PIN code 400 017 flagged. Training data associates this area with higher historical defaults. Rahul's score is downgraded. He is not told this variable exists.
Step 3 Older Android = risk proxy. Device age correlated with repayment in training data. Another downgrade.
Step 4 9 PM application flagged. Late-night applications correlate with distress-borrowing. Rahul applied then because it was his free time.
Michael
Approved
11.2% interest. He assumes his credit score was the reason.
Rahul
Rejected
"Does not meet our current lending criteria." No reason. No appeal. Reapply in 6 months.
Rahul cannot contest the PIN code signal. He was never told it existed. How would you know if an AI decision about you was biased?