Three Takeaways From the GSE Collateral Symposium on AI and Computer Vision

By Vin Vomero, CEO, FoxyAI

This week I was invited to speak at Fannie Mae and Freddie Mac’s Collateral Symposium on AI and Computer Vision, an invite-only event that brought together a concentrated group of people from the GSE’s, and the appraisal, valuation, MLS, lending, and property data sectors.

I was on the Computer Vision 101 panel, but the more interesting part for me was hearing how the rest of the industry is thinking about where we are headed as an industry.

Three things stood out.

1. We’re moving from data plus images to data from images

Freddie Mac’s Danny Wiley summed it up well: we’re moving from “data plus images” to “data from images.”

Historically, photos provided supporting evidence. Someone inspected a property, entered the data, and attached images to back it up.

Computer vision changes that. The images themselves can now generate property data, such as condition, quality, features, measurements, missing-photo checks, and other structured information.

That is a big part of what we do at FoxyAI. We use property imagery to turn what was previously visual, manual information into structured data that can be used in valuation, underwriting, servicing, and other workflows.

The important shift is not just that AI can “analyze photos.” It is that imagery is becoming a machine-readable property data source.

And that creates a much bigger opportunity: collect better imagery once, extract more information from it, and reuse that data throughout the property lifecycle.

2. Image manipulation and fraud are becoming a much bigger problem

This was probably my biggest takeaway from the day.

We’ve had digital photos for decades, and photo manipulation is not new. The difference is that it used to require Photoshop, time, and at least some artistic skill.

Generative AI has changed the landscape completely.

Anyone can now create or alter a photorealistic image in seconds. As evidenced by examples from John Holbrook throughout the day, it’s now virtually impossible to tell the difference between real photos from those that have been altered. 

That matters a lot in mortgage and real estate, where photos can influence property condition, appraisal, underwriting, servicing, insurance, and collateral decisions.

A major theme of the discussion was that trying to detect manipulation after the fact is a difficult game of cat and mouse to win. If authenticity really matters, the strongest control is establishing trust at the point of capture: where the image came from, when it was taken, what device captured it, and whether it was altered before entering the workflow.

That is relevant to us at FoxyAI too. We already analyze large volumes of property imagery for condition, quality, features, and damage. Increasingly, customers are asking a second question before any of that:

Can I trust the image in the first place?

I think image provenance and fraud detection are going to become their own layer of the property data stack, not just another computer vision feature.

3. The bigger opportunity is redesigning the property data workflow

The MLS discussion surfaced another important issue: listing photos are generally created to market a house, not to serve as decision-grade collateral evidence.

But appraisers, lenders, valuation models, and AI systems increasingly reuse those same images.

That juxtaposition creates a mismatch exacerbates the issue. 

A brighter room or improved sky may be harmless marketing. Removing damage, changing materials, or making a comparable property look materially better can affect downstream decisions.

The more interesting opportunity is to stop relying entirely on data that was created for another purpose.

Someone is already going to the property and taking photos. What if that same visit could also produce authenticated imagery, measurements, floor plans, condition information, and standardized property data that could move through the rest of the transaction?

That starts to eliminate repeat visits, duplicate data entry, and inconsistent property information.

And that is where I think computer vision gets much more interesting.

Not as a better way to analyze a pile of photos after the fact, but as infrastructure for collecting property data correctly once and making it useful across the transaction.

Summary

The technology is getting better very quickly. The harder questions now are around trust, standards, workflow design, and deciding where humans still need to make the call – this  was the broader message I took away from the symposium.