The 3 Forensic Checks That Expose a Deepfake Your Eyes Will Never Catch

A single round of J.P.E.G. compression can erase nearly a third of the evidence that proves a face is fake. Not the face itself. Not the expression or the skin tone. The invisible mathematical fingerprints hiding in the pixels — gone, just from uploading to social media.

That should unsettle you, whether you analyze

That should unsettle you, whether you analyze digital evidence for a living or you just scrolled past a video this morning and thought, "that looks real enough." Because the deepfakes arriving right now don't fail the eye test anymore. According to researchers studying generative adversarial networks — G.A.N.s, the A.I. systems behind synthetic faces — the latest models produce skin texture, lighting, even eye reflections that a non-expert will almost always accept as genuine. If that feels scary, it should. But there's a structured way to fight back, and it doesn't depend on how sharp your eyes are. It depends on understanding three forensic checks that catch what human vision can't. So what are those checks, and why do they work when our own perception doesn't?

The first check targets something called artifact analysis — and the two places it matters most are the mouth and the eyes. Not because those features look wrong to you. They probably look fine. The problem is deeper than appearance. When a deepfake algorithm generates a face, it handles complex regions like lips and irises differently from flatter areas like cheeks or foreheads. Researchers categorize the result as Face Inconsistency Artifacts. Basically, the intricate parts of the face and the simpler surrounding areas don't degrade at the same rate. The mouth reveals lip-sync mismatches at the pixel level. The eyes produce reflections that real corneas generate naturally but synthetic ones can't quite replicate. You won't notice that mismatch on a phone screen. But a forensic tool measuring how pixel quality shifts between the eye region and the cheek beside it — that catches the seam every time.

And there's a second category of artifact that's even more universal. Every G.A.N.-generated face goes through a step called up-sampling — the decoder inside the network scales the image up to full resolution. That scaling process leaves behind a repeating checkerboard pattern in the frequency domain. The frequency domain is just a different way of looking at an image — instead of seeing colors and shapes, you see the mathematical wave patterns underneath. Imagine holding a counterfeit bill under a blacklight. A real bill has security fibers woven in during manufacturing. The counterfeit might look identical in daylight, but under that blacklight, the fibers are missing because the forger never had access to the original manufacturing process. Frequency analysis does the same thing for faces. The checkerboard pattern from up-sampling is a manufacturing trace the G.A.N. can't avoid leaving behind. For someone reviewing evidence, that's a reproducible signal. For the rest of us, it means a photo that fools every person in the room can still betray itself mathematically.

Why not just run every suspicious image through a

So why not just run every suspicious image through a detection tool and call it done? Because of compression — and this is the second forensic check: source tracing. When a video gets uploaded to a social platform, the platform compresses it. That compression introduces its own artifacts. According to academic research on deepfake forensics, some compression artifacts are nearly indistinguishable from the artifacts a deepfake creates through affine transformation — that's when a synthetic face is mapped onto a real person's head. The resolution mismatches from that mapping look a lot like the resolution mismatches from aggressive video compression. A single pass of J.P.E.G. compression can mask up to thirty-two percent of detectable artifacts. That's why source verification matters so much. Investigators need to trace the file back — check its metadata, its compression history, whether it passed through platforms that strip forensic data. And for anyone sharing a video that seems too outrageous to be true, this is worth knowing: the very act of downloading and re-uploading that clip may have erased the proof that it was fabricated.

The third check is cross-image consistency, and it exploits something most people don't realize about G.A.N.s. These networks can only produce a limited range of brightness values. They don't generate truly saturated highlights or deeply under-exposed shadows. A real camera capturing a real scene records the full spectrum of light — blown-out whites, crushed blacks, everything in between. A G.A.N. operates in a narrower band. That creates a specific lighting signature. If you compare multiple images allegedly from the same scene, a real set will show natural variation in shadow depth and highlight intensity. A synthetic set will cluster in the middle of that range. What does that mean practically? It means someone examining a batch of suspect photos can check whether the highlights and shadows behave the way a physical camera would produce them. And if you've ever wondered why a deepfake video sometimes feels slightly "flat" even though you can't articulate why — this limited dynamic range is part of the answer.

One more thing that ties all three checks together. Researchers have found that manipulation artifacts are local in nature. There's no big, obvious difference between a real image and a manipulated one when you look at the whole picture. But in the specific regions where the algorithm did its work — the eyes, the mouth, the boundary where the synthetic face meets the original — the same mathematical fingerprint repeats. And it repeats across every image made by the same method. That consistency is what lets analysts link multiple suspect images back to a single source tool.