A Deepfake Fooled a Notary on a Live Call. The Ears Gave It Away.

Here's a number that should genuinely unsettle you: 60% of people are confident they can spot a deepfake. And in practice, only 0.1% of them actually can. That's not a rounding error — that's a chasm. And nowhere is that chasm more expensive than in real estate, where a misplaced confidence in human pattern recognition is currently costing the industry hundreds of millions of dollars a year.

In Maryland, a fraudster nearly walked away with roughly $100,000 from a vacant land transaction by impersonating the property owner during a live video notarization. Not a pre-recorded clip. A live call. The deepfake was convincing enough to pass notary observation in real time — which should make every title officer and closing agent in the country deeply uncomfortable. What caught it wasn't a sharp-eyed human. It was detection software flagging spatial inconsistencies that no human reviewer would have noticed on a video call.

That case is a preview of where fraud is going. And understanding why software catches what humans miss requires understanding how faces actually work — as geometry, not just appearance.

---

Why Real Estate Is Wearing a Target

Remote closings were supposed to make transactions smoother. They did. They also eliminated the one friction point that made identity fraud genuinely hard: standing in a room with another human being. According to HousingWire, FBI data shows real estate fraud generated 12,368 complaints and $275 million in losses in 2025 alone. Deepfake-enabled schemes are a growing slice of that number, and the math is straightforward from a criminal's perspective: a single convincing deepfake identity video can unlock a transaction worth ten, twenty, fifty times more than it cost to produce.

Deepfake-related scams are growing at 40% year over year, per Entrust's 2026 Identity Fraud Report. That growth rate is outpacing the adoption of detection tools, which means the gap between what attackers can produce and what investigators can catch is widening — right now, in active transactions. The industry isn't losing ground slowly. It's losing it fast.

And here's something investigators often underestimate: the video call isn't even the hardest part of the fraud to construct. The Voice of San Francisco reports that just three seconds of recorded audio can generate an 85% voice match for a synthetic clone — though most operators target at least 30 seconds for a convincing result. Voicemails. Listing videos. A recorded introduction at an open house. Every fragment of audio a real estate professional puts online is training data for their own impersonation. This article is part of a series — start with Deepfake Laws Biometric Standards Gap Investigators.

This is why the video evidence — the ID selfie, the notarization call, the authorization clip — becomes the weak link worth attacking. And it's where facial comparison earns its place in an investigator's toolkit.

---

The Myth of "Just Watch Carefully"

Let's correct something, with some empathy for why it's wrong.

Most people's mental model of a deepfake comes from 2017–2019, when the technology was genuinely crude. Faces flickered. Edges blurred. Blinking looked mechanical. The media covered these failures — and in doing so, accidentally trained millions of people to associate "deepfake" with "obvious fake." If you've seen a bad deepfake and caught it instantly, your brain filed that experience under I know what to look for.

But that instinct is now working against you. As Florida Realtors documents, modern generative adversarial networks trained on thousands of hours of video now produce eye blinks that follow statistically normal human distributions. Facial hair renders with follicle-level detail. The real estate fraudster's deepfake in 2026 doesn't have a telltale shimmer. It doesn't stutter. The notary on that Maryland call wasn't negligent — they were up against technology that was specifically designed to pass human inspection.

"Synthetic media becomes more realistic and spotting it with the naked eye is harder than most people think — even experienced professionals are vulnerable to high-quality deepfakes." — Florida Realtors, floridarealtors.org

The artifacts that do exist in modern deepfakes are geometric, not aesthetic. When a hand passes in front of a deepfake face, it often warps. When a mouth moves extensively, its horizontal position relative to the nose-tip-to-chin axis can drift — subtly, imperceptibly to the eye, but measurably. Facial hair at the jawline where the synthetic face meets the real neck can show micro-inconsistencies. These aren't things you see. They're things you measure.

---

468 Points of Truth

Here's how structured facial comparison actually works — and why it catches what eyes don't.

A facial landmark analysis framework maps 468 distinct points across a human face: the corners of each eye, the precise peaks of the cupid's bow, the attachment point of each earlobe, the angle where the jawline meets the neck. Each point has a coordinate. And between any two points, you can calculate a distance — either as a straight Euclidean measurement ("how far apart in millimeters?") or as a Geodesic distance that accounts for the face's three-dimensional curvature. Previously in this series: A Fake Cfo Stole 25 6m The Real Victim Is Your Evidence Proc.

The investigative power isn't in any single measurement. It's in the ratios. The ratio of the distance between a person's pupils to the distance between their cheekbones doesn't change when they smile, age five years, or grow a beard. It's a biometric constant baked into bone structure. Deepfake generation tools — which work by transplanting one person's facial surface texture onto another person's head — cannot alter these ratios without breaking the realism of the output. They're constrained by the skull geometry underneath the synthetic skin.

Think of it like this: imagine comparing two architectural blueprints — one authentic, one forged. A skilled forger can replicate the building's overall appearance convincingly, adding realistic shadows and accurate-looking dimensions. But overlay both blueprints on a grid and measure fifty precise distances — corner A to corner B, wall junction C to window center D — and the forgery will have six distances that simply don't match the original. The forger copied the look, not the geometry. Deepfakes work exactly the same way. They copy the appearance of a face without preserving the geometric invariants that make it that specific person's face.

At CaraComp, this is the methodology that separates a gut-check from a defensible finding — running structured multi-landmark comparison across multiple reference images, calculating Euclidean distance ratios at key biometric anchors (eye-to-ear span, nasal bridge to ear canal, mandible angle), and documenting where a questioned video's face diverges from the known reference. The ears, specifically, are often where deepfakes fail first. They're structurally complex, asymmetric between individuals, and geometrically difficult for face-swap algorithms to render accurately when the head turns even slightly.

"Check the ears" sounds almost too simple. It isn't. It's where the geometry breaks.

---

What an Investigator Actually Does With This

In a real estate fraud scenario, an investigator doesn't just compare the video call to an ID photo and call it done. The methodology requires multiple reference images — ideally pulled from sources the suspected fraudster couldn't have anticipated or controlled. DMV records. Older social media posts predating the fraud attempt. Professional license photos. Court filings. Each image contributes additional landmark data that either confirms or contradicts the questioned video.

The comparison is then documented as a Euclidean distance analysis across those landmarks — not as "these faces look different" but as "the measured ratio of interpupillary distance to bizygomatic width deviates by X standard deviations from the reference set, consistent with a face-swap artifact rather than natural photographic variation." That's the difference between an observation and evidence. Up next: The Cop Who Made 3 000 Deepfakes Exposed A Bigger Problem Th.

According to Stewart Title, forensic detection methods applied to real estate fraud cases have identified manipulation in owner authorization videos, manipulated listing images, and remote closing verification clips — exactly the materials that most title companies currently treat as reliable proof of identity.

The window for stopping fraud is narrow. Once wire transfer instructions are followed, recovery is statistically unlikely. Which means the facial comparison needs to happen before the money moves — as part of transaction due diligence, not post-incident forensics.

The Maryland transaction was caught. But it was caught by software, not by the trained professional on the other end of the call. That professional had years of experience, sharp eyes, and complete confidence in their ability to assess what they were seeing. And the deepfake still passed them.

So here's the question worth sitting with: if you suspected a deepfake in a property transaction right now, what's the first piece of visual evidence you'd try to verify — the government ID photo, the selfie verification video, or the social media images the "buyer" used to establish their identity? The answer matters, because each of those has a different vulnerability profile, different availability of reference landmarks, and a very different likelihood of catching a sophisticated synthetic face before it costs someone their property.

The ears don't lie. But you have to know to look at them — and know how to measure what you're seeing.