AI-generated fraud: synthetic borrowers & deepfake collateral

Generative AI broke the assumptions private lending was built on.

Underwriting in private real estate has always rested on a quiet assumption: it is hard to fake a photo of a job site, hard to fabricate a bank statement, hard to invent a borrower. None of those are hard anymore. Any underwriter operating on 2022 assumptions is exposed.

What changed in 18 months

Image models can produce photo-realistic construction progress at any stage of completion. Document models can produce bank statements that pass a casual review. Voice clones can pass a callback. Identity models can build a synthetic borrower with a credit footprint that looks real for the first 90 days.

Detection is necessary but not sufficient

Classifier-only defenses lose ground every model release. The durable defense is provenance: capture evidence on a controlled surface, sign it at the moment of capture, and store the chain. Detection becomes the second line, not the first.

The lender response

Treat any borrower-submitted artifact as untrusted by default. Require capture through a controlled surface for collateral evidence. Run AI-generation classifiers on every document at intake. Score the file, not the page. Make it easy for a human to override with a logged reason — that audit trail is what protects the lender later.

You cannot out-classify a model release cycle. You can out-design it with provenance — capture, sign, store, and verdict at the moment of capture.
— Zack Farris
Source for this story?

I am on record on this topic. Reporters, podcast producers, and panel organizers can reach me directly through the press kit.