Deepfakes Are Flooding Schools. Here's the Forensic Trick That Actually Catches Them.

Here's a number that should stop you mid-scroll: reports of AI-generated child sexual abuse images submitted to the National Center for Missing and Exploited Children jumped from 4,700 in 2023 to 440,000 in just the first six months of 2025. That's not a trend line. That's a vertical wall. And the place where a growing share of these images first circulate isn't the dark web — it's a school group chat.

When one of those images lands in a principal's inbox — forwarded by a panicked parent at 7 a.m. — the clock starts immediately. Parents want answers in hours. Police may need evidence within 48 hours. And the administrator standing in the middle of it all is almost certainly working without a protocol, a trained investigator, or any clear idea of what "proof" even looks like in this context.

That's the real story here. Not just that deepfakes are circulating in schools — they are, at scale — but that the investigation of a deepfake is a specific, learnable forensic process. One that most schools have never been taught, and one that starts with a question most people get wrong.

The Wrong First Question

When someone hands you a suspicious image, the instinct is to ask: "Is this real?" That feels like the right place to start. It isn't. And understanding why changes everything about how you approach the investigation.

Human beings are genuinely terrible at spotting deepfakes. According to data compiled by SQ Magazine, people successfully identify high-quality deepfake videos only about 24.5% of the time. For images, accuracy climbs to around 62% — which sounds better until you realize that's barely better than a coin flip with extra steps. In mixed tests across modalities, only 0.1% of participants could reliably detect fakes. Not 1%. Point-one percent.

So when a staff member looks at a suspicious image and says "I can't tell if this is fake," they're not failing at their job. They're performing exactly as human visual perception is designed to perform — which is to say, nowhere near well enough for this task. This article is part of a series — start with That 95 Face Match Scammers Built The Other 3 Layers To Fool.

The misconception that follows is the dangerous one: if I can't see proof, there is no proof. Schools make this mistake constantly. The inability to spot a deepfake by eye doesn't mean the investigation is over. It means the investigation needs to actually begin — using methods that don't rely on human intuition at all.

What the Forensics Actually Look At

Deepfake detection, when done properly, works by examining specific facial regions for inconsistencies that neural network blending almost always introduces. The field calls these regions facial landmarks — and there are dozens of them, including the inner corners of the eyes, the tip and bridge of the nose, the corners of the mouth, and the jaw boundary where a synthesized face meets its background.

Peer-reviewed research published in MDPI's Information journal demonstrates that fusing eye, nose, and mouth landmark data produces detection accuracy with an AUC of 0.875 — meaning the model correctly distinguishes real from fake about 87.5% of the time on datasets featuring unnatural eye movements alone. A separate model published on Preprints.org using temporal convolutional networks reached an F1 score of 0.917 when analyzing eye-nose fusion patterns across frames.

What makes these landmarks so revealing? When an AI generates or blends a face, it has to make thousands of micro-decisions about spatial relationships — how far are the inner eye corners from each other? How does that distance change as the head turns? Does the shadow under the nose move consistently with the light source implied by the background? Human faces follow physics. AI-generated faces follow training data, and training data has gaps.

Think of it like examining a forged signature on a check. Eyeballing it might make you suspicious — something feels off. But proving forgery requires a forensic document examiner to identify specific inconsistencies: pressure variations, stroke angles, the slight tremor a forger introduces when trying to slow down and be precise. The forgery isn't caught by vibes. It's caught by measurement. Deepfake investigation works the same way — and the measurements are in the landmarks.

"Even trained professionals are struggling, and some journalists admit they can no longer reliably identify deepfakes without using forensic tools." — Daon, Next-Gen Deepfake Detection Report

Beyond landmark geometry, investigators look at three additional layers. First: temporal consistency — in video, does lip movement sync cleanly with audio across different head poses, or does the sync break when the subject turns even slightly? Second: texture boundaries — at the chin-jaw edge where synthesized skin meets the original background, neural blending often leaves a telltale softness or color temperature mismatch. Third: iris reflections — real eyes reflect a consistent light source. GAN-generated eyes frequently show reflections that contradict the ambient lighting in the rest of the image. Previously in this series: Uk Scanned 1 7m Faces Seven Regulators Cant Agree On The Rul.

What Happens When Schools Skip This Step

In Iowa, four boys were charged in juvenile court after using AI tools to generate fake nude images of 44 girls, sourced from ordinary social media photos, according to reporting aggregated by Townhall. The same reporting notes that across roughly 90 schools globally, more than 600 students have been affected by similar incidents. In a Louisiana middle school, AI-generated images spread through the student body so quickly that — before any investigation was complete — one of the victims was expelled for physically confronting a boy she suspected of creating the images.

Read that again. The victim was expelled. Because the school had no forensic framework, no way to rapidly document what the images were or where they came from, the response defaulted to managing chaos rather than establishing facts.

According to a RAND Corporation survey, 13% of principals reported deepfake incidents during the 2023–2024 and 2024–2025 school years — with 22% of high school principals and 20% of middle school principals reporting cases. Yet only 23% of schools updated their policies to include any specific language about AI misuse. The other 77% are improvising.

The National Education Association reports that between 40 and 50 percent of students are already aware of deepfakes circulating at their school. Fabrication takes seconds using free apps. A single photo from a public social media profile is enough raw material. The asymmetry is brutal: creating a deepfake is trivially easy; investigating it properly requires trained methodology that most schools simply don't have.

From Suspicious Image to Documented Evidence

Here's what systematic visual analysis actually gives you — and why it matters beyond the technical result. At CaraComp, the work of image comparison isn't just about getting a confidence score. It's about producing documentation that a parent, a police detective, or a district attorney can follow. That requires a specific kind of output: not "we think this is fake," but "the inner canthal distance — the gap between the inner corners of the eyes — shifts by 4.3 pixels across equivalent frames in a way inconsistent with natural head movement, and the chin boundary shows luminance artifacts at 2–4Hz consistent with GAN blending."

That sentence transforms a rumor into evidence. And it changes the conversation in a principal's office from "we can't really prove anything" to "here is what we found and here is what it means." Up next: Retail Facial Recognition Watchlists No Appeals Process.

The multi-step process matters too. Landmark analysis is the starting point — not the whole investigation. Source-chain verification (where did the image first appear, and on which platform?) and metadata examination (what device, what timestamp, what editing software fingerprint does the file carry?) work together with facial forensics to build a complete picture. Research covered in depth by Springer Nature identifies lighting inconsistency and shadow analysis as additional forensic layers — because generated faces are often composited onto backgrounds with incompatible light sources, and that inconsistency is measurable.

No single signal proves manipulation. The power is in the convergence: when the landmark geometry is off, the metadata is missing, and the lighting physics don't add up, you have documentation that holds.

The deeper question worth sitting with: if a school brought you one suspicious image right now, which would you trust first — facial feature consistency, metadata, or source-chain analysis? Most people instinctively reach for metadata (it feels objective, like a timestamp on a receipt). But metadata is trivially stripped or spoofed. Facial landmark inconsistency, analyzed frame by frame, is far harder to fake — because the neural network that generated the face didn't know it would be scrutinized at the pixel level.

That's the insight worth keeping. The best evidence in a deepfake case is usually hiding in the face itself, in the 3.7-pixel gap between where the eyes are and where they should be. Schools don't need to become forensics labs overnight. But understanding that this evidence exists — and that it's findable — is the prerequisite for everything that comes next.