- A $2.5M settlement just proved what we've been saying:
Most AI lending systems aren't biased because of bad intentions.
They're biased because of lazy architecture.
The era of "the algorithm decided" is over. 🧵 - Earnest Operations used a school's Cohort Default Rate to score borrowers.
Sounds neutral. It's not.
HBCUs have higher CDRs due to systemic underfunding — not because their graduates are riskier.
The model penalized individuals for a group-level historical artifact. - Worse: Earnest had "knockout rules" that auto-denied anyone without a green card.
No human review. No override path. Just a hard gate coded into the system.
The AG called it what it was — unfair and deceptive.
Then there's Navy Federal Credit Union. - 77% approval rate for white mortgage applicants.
48.5% for Black applicants.
Researchers controlled for income, DTI, property value, and more.
Black applicants were still 2x more likely to be denied. The gap couldn't be explained away. - Both cases share a root cause: shallow AI architecture.
LLM wrappers pass data to a foundation model and call it "AI-powered underwriting."
No domain-specific guardrails. No fairness constraints. No causal reasoning.
Just a probability engine wearing a compliance costume. - The CFPB has made it explicit: "the algorithm decided" is not a legal defense.
Adverse action notices must cite accurate, specific reasons.
If your system can't explain WHY it denied someone in auditable detail, you're already non-compliant. - This is why we built around Deep AI — not wrappers.
Deterministic rule engines for hard compliance checks.
Gradient boosted models for interpretable credit scoring.
Fine-tuned LLMs only for document analysis, grounded by RAG.
Each layer does what it's actually good at. - Fairness isn't a checkbox. It's math.
Adversarial debiasing. SHAP-based explainability. Real-time disparate impact monitoring.
If your model's fairness ratio drifts below threshold, the system flags it before a regulator does. - The uncomfortable truth: most fintechs haven't even inventoried which of their systems qualify as "models" under SR 11-7.
No inventory means no validation. No validation means no defense.
We think the industry is splitting in two: - Institutions building defensible AI architectures from the ground up.
And institutions waiting for their own $2.5M lesson.
Which side is your org on? #AICompliance #FairLending - We wrote the full breakdown — proxy discrimination, fairness engineering, the Deep AI framework — in our latest whitepaper:
https://veriprajna.com/whitepapers/algorithmic-accountability-deep-ai-financial-services-enforcement