Banks Confront Deepfake Borrowers in Automated Lending

PYMNTS reported on June 10, 2026 that fraud operators are combining deepfake video, cloned voices, fabricated employment histories and AI-generated financial behavior into single engineered "synthetic borrower" personas designed to pass automated onboarding and underwriting, then disappear once loans are funded. James Mirfin, global head of risk and security intelligence solutions at Visa, and Adam Hiatt, vice president of fraud strategy at fraud-prevention vendor Spreedly, told PYMNTS that AI systems can now replicate the behavioral consistency that human reviewers and anomaly detectors are built to check for; Hiatt called it "an arms race."

Traditional fraud detection leans on identifying statistical outliers: sudden credit-activity spikes, inconsistent identity data, or unusual onboarding behavior. Synthetic borrowers invert that assumption by engineering personas to fall within normal ranges across video, audio, documents and simulated financial behavior at once, which degrades anomaly-based classifiers and can bias downstream credit-risk scores as this data enters training pipelines. PYMNTS Intelligence research cited in the piece found fraud now spans the full customer lifecycle rather than just onboarding, and that 77% of credit unions surveyed reported unauthorized network access in the past year.

The multi-modal deepfake angle sits inside a faster-growing synthetic-identity-fraud problem. Equifax has reported synthetic-identity losses rose 50% between 2022 and 2023, with industry estimates it cites putting current annual U.S. losses at $20 billion to $40 billion; Deloitte projects losses could reach $23 billion annually by 2030, and TransUnion measured $3.3 billion in synthetic-identity exposure for lenders in 2024. Point Predictive, an auto-lending fraud vendor, separately reported, in its own vendor-published figures, that AI- and deepfake-related mentions on fraud-focused Telegram and dark-web channels rose from about 47,000 to more than 350,000 between 2023 and 2024, and put 2026 auto-lending fraud exposure at $10.4 billion. In response, Equifax launched a Synthetic Identity Risk tool and a Credit Abuse Risk model in January 2026 that separate identity verification from creditworthiness scoring.

Anomaly detection alone is no longer sufficient once attackers can target the statistical center of a distribution rather than its edges. That argues for layering in provenance and liveness signals that flag AI-generated media and documents directly, separating identity-verification models from credit-decisioning models the way Equifax has, and treating cross-institution data-sharing as a core control, since no single lender can see loan-stacking happening across competitors.

Watch for lenders or vendors to publish fraud-loss or vintage-default figures tied specifically to AI-generated synthetic borrowers, since most numbers so far describe synthetic-identity fraud broadly rather than the deepfake-specific subset; for wider adoption of multi-modal liveness and provenance detection; and for more bureaus or regulators following Equifax's lead in separating identity verification from credit scoring.