AI in mortgage risk: promise vs. perils
Lenders are rapidly adopting AI for underwriting and portfolio analytics, but regulators and academics warn 'black‑box' models risk operational, reputational, and systemic issues if adopted without rigorous oversight. The debate is intensifying: AI can boost predictive power, yet unchecked models could amplify correlated losses across lenders. (gsb.stanford.edu) (nationalmortgagenews.com)
Tidalwave and Columbia University’s DAPLab released the first public benchmark for mortgage-origination AI on March 16–17, 2026, finding Tidalwave’s mortgage‑trained SOLO scored as high as 95% on yes/no underwriting compliance checks while Anthropic’s Claude 4.5 scored 42% on the same tasks. (daplab.cs.columbia.edu) The U.S. Federal Housing Finance Agency and GSEs confirmed they terminated use of Anthropic products in early March 2026 following a federal directive, forcing immediate vendor‑risk re‑assessments for lenders who integrate Claude‑based tools into origination or servicing workflows. (marketscreener.com) Canada’s OSFI and the Financial Consumer Agency flagged AI uses and risks in a joint report published Sept. 24, 2024 and circulated a voluntary AI/quantum‑readiness questionnaire in December 2023 as part of broader efforts to tighten model governance for federally regulated institutions. (publications.gc.ca) OSFI has signalled it will expand model‑risk expectations via updates to Guideline E‑23 (drafted Q4 2023) that explicitly call out explainability, governance and testing requirements for ML/AI models used in material decisions. (osler.com) TD Bank began rolling out AI‑driven instant mortgage and HELOC pre‑approvals in May 2024 through its Layer 6 capabilities, and Royal Bank of Canada closed an acquisition of mortgage fintech Pinch Financial on March 12, 2026 to scale automated borrower verification across origination channels. (stories.td.com) Consultancies and central‑bank analysts warn concentration and valuation risks tied to AI could transmit correlated losses if models or vendors fail; Oliver Wyman’s January 2026 analysis models equity and hybrid bubble scenarios that could amplify market stress if AI exposures are widespread. (oliverwyman.com)
