How AI’s Black-Box Decisions Threaten the Mortgage Repurchase Framework

How Non-Deterministic AI Breaks the Mortgage Trust Chain

The mortgage industry’s centuries-old architecture of trust – a delegated chain from government-sponsored enterprises to lenders, loan officers and borrowers – is buckling under the weight of non-deterministic artificial intelligence, warns Marvin Chang, Associate Director of the FinTech program at Duke University’s Pratt School of Engineering. The problem is rooted in the repurchase and warranty framework that has long held the system together: when a loan goes bad, the original decision-making can be examined, fault assigned and repurchase demands issued. The Global Financial Crisis showed what happens when documentation lags the delegation of trust, triggering billions in settlements and permanent changes in lender strategy.

Today, AI systems that drive underwriting, fraud detection, automated valuation models and income verification are fundamentally different from the rules-based engines they replace. They produce outputs with a reasoning path that is not preserved in a form audits can reconstruct, and – critically – the same inputs can yield different outputs on different days. “The system may not know why it said no,” Chang writes. “Which means you can’t tell the borrower why you said no. Which means you are exposed – to the borrower, to the regulator and to the repurchase framework.”

The exposure is magnified because modern mortgage origination runs on a layered stack of third-party AI vendors. A single loan may pass through point-of-sale systems, AVMs, fraud detectors and income verification tools, each making opaque judgment calls that feed the next. No single lender has full visibility into that chain, and when a loan fails, the question of which system introduced the error – and whether it can be audited – may be unanswerable. This structural vulnerability cannot be patched with more paperwork; it requires an architectural response embedded in how systems are built, Chang argues.

Where the Liability Lies: Multi-Vendor AI’s Accountability Gap

Why Non-Deterministic AI Escapes Traditional Audits

Rules-based decision engines leave a deterministic trail: if input X, then output Y. Regulators and repurchase counterparties can re-run the logic and verify the decision. Non-deterministic AI weighs inputs probabilistically, generating a conclusion without a linear, reproducible record. The model’s internal reasoning may be opaque even to its own developers, making it impossible to prove, after the fact, that a loan decision was sound. This undermines the core assumption of the rep and warrant structure – that the lending process is demonstrable.

The Multi-Vendor Problem: No Single Owner of the Decision

Chang’s central insight is that the true exposure lives at the interfaces between AI systems, not within any single model. A decision distributed across four or five black-box vendors cannot be reconstructed by examining each vendor in isolation. Unlike traditional vendor relationships, where logic could be contractually specified and checked, AI vendor outputs are inherently variable. The vendor may not be able to explain the reasoning any more than the lender can, leaving a governance black hole that sits outside the compliance programs the industry spent years building. When a loan sours, liability becomes a finger-pointing exercise with no clear answer.

What a Repurchase Wave Could Look Like

If confidence in accountability erodes, capital becomes more cautious, more expensive and slower to deploy – a dynamic Chang compares to Argentine property transactions conducted in U.S. dollars because the peso cannot be trusted. A future enforcement action or repurchase wave triggered by AI-driven defects would force lenders to absorb losses they cannot trace, and the GSEs may retreat from delegating underwriting authority to systems they cannot examine. The first lender that can demonstrate full workflow-level auditability across its AI stack stands to gain a significant competitive advantage in capital access and counterparty trust.

MISMO’s FRAME Initiative as a First Step

Industry recognition of the problem is growing. The Mortgage Industry Standards Maintenance Organization (MISMO) has launched FRAME (Framework for Responsible AI in the Mortgage Ecosystem), an effort to bring governance to AI in mortgage. Chang acknowledges this as a genuine step but insists the next phase must address what happens between systems, not merely within them. Governance in a multi-vendor AI environment must be organized around the decisioning workflow – the full sequence from input to output across every system that touched the loan.

Three Shifts Lenders Must Make to Rebuild Auditable AI Workflows

Chang outlines three practical shifts that mortgage lenders – particularly chief risk officers, heads of capital markets and vendor management teams – should pursue now:

• Rewrite vendor contracts to document cross-system interactions. Current agreements typically specify only a single model’s output. New contracts should require each vendor to detail how its AI outputs are consumed by downstream systems and under what conditions an output may change, creating a contractual map of the decision chain.
• Implement workflow-level logging owned by the lender. Instead of relying on vendor-side records after a problem emerges, lenders should build their own logging infrastructure that captures every AI-generated judgment call, the system of origin and the sequence of inputs and outputs across the full origination stack. This makes the question “where was the decision made?” answerable from the lender’s own data.
• Design auditability into the architecture, not as a post-hoc layer. Risk teams should require AI models to produce deterministic summaries or a reproducible audit artifact alongside their non-deterministic output, enabling a reconstruction of the decision without reverse-engineering the model itself. This approach, which Chang calls an architectural response, shifts the burden from forensic reconstruction to upfront design.

Lenders that move first on these fronts will be better positioned when the first major AI-related repurchase wave arrives. They will be the ones that capital providers, the GSEs and the courts trust to extend delegation to – a structural advantage that could define the next cycle of mortgage finance.