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

Here's something that should make you stop scrolling: a face can be mathematically perfect and still be completely fake. Not "slightly off" fake. Not "uncanny valley" fake. Indistinguishable to the human eye fake — and the only way to know the difference is to stop looking at the face and start looking at what the algorithm left behind.

When a viral face image drops on social media — a celebrity, a political figure, someone suddenly famous for all the wrong reasons — the public's first instinct is to assess it visually. Does the skin look real? Are the eyes tracking right? Does the lighting make sense? These are reasonable questions. They're also almost entirely the wrong ones. The real evidence of manipulation isn't in what you see. It's in what the generation process couldn't help but leave behind.

Why Your Eyes Are the Worst Tool for This Job

Modern generative adversarial networks — GANs, the architecture behind most deepfake face synthesis — are explicitly trained to defeat visual inspection. The entire adversarial training loop is, at its core, an ongoing war between a generator trying to fool human perception and a discriminator trying to catch it. After millions of training iterations, the generator gets extremely good at producing faces that register as "real" to anyone glancing at a screen.

But here's what the generator can't escape: the mathematical steps it takes to build that face leave structural traces in the pixel data that no amount of visual polish can hide. Researchers studying deepfake detection have identified two fundamental categories of these traces. The first are Face Inconsistency Artifacts (FIA) — inconsistencies that emerge specifically when the algorithm tries to synthesize intricate facial details like pores, lip texture, and eyelash structure, creating mismatches between those complex zones and the smoother surrounding regions. The second are Up-Sampling Artifacts (USA) — patterns baked into the image during the generator's decoding process, present in essentially every GAN-produced face image regardless of which model created it.

The USA category is particularly important because it's universal. According to arXiv research on generalizable deepfake detection, up-sampling artifacts appear across all existing deepfake generation methods — meaning even as GAN architectures evolve, this class of evidence persists. Investigators don't need to know which specific tool generated a suspect image. The artifact class is consistent regardless.

Check 1: Artifact Review — Where to Actually Look

Not every pixel in a deepfake image holds equal investigative value. Detection systems have consistently converged on two facial zones as the highest-yield targets: the mouth and the eyes. This article is part of a series — start with Federal Judges Just Gutted The Its Real Defense And Investig.

Why those two? The mouth is where lip-sync generation creates the most computational strain — the algorithm has to model depth, wetness, shadow under the upper lip, the way teeth catch light differently from gum tissue. Small failures cluster here. The eyes are worse for deepfakes in a different way: real eyes produce corneal specular highlights — those small white reflections from ambient light sources — that follow physically consistent rules. GAN-generated irises frequently show reflections that don't correspond to any light source in the rest of the image, or produce pupil shapes that shift subtly between frames in ways biological eyes don't.

Beyond these specific zones, comprehensive deepfake forensics research shows that synthetic manipulations disrupt texture consistency in ways that are visible in both the spatial domain (how pixels relate to their neighbors) and the frequency domain (how those relationships look when mathematically transformed). This is where the analogy earns its keep.

Think of a deepfake like a counterfeit document. A skilled forger can replicate everything you see under normal light — ink color, paper texture, signature style. But under a blacklight, a genuine document's security fibers create patterns the counterfeit cannot replicate, because those fibers were embedded during the original manufacturing process. GAN-generated images have the same problem in the frequency domain: University of Missouri research on spatial and spectral deepfake detection identifies checkerboard artifacts in frequency-analyzed GAN images as a direct result of the up-sampling process — a structural signature that looks invisible on screen but is mathematically undeniable when you transform the image into the frequency domain.

There's one more artifact worth knowing: GAN generators produce images with a constrained intensity range. Real cameras capture the full spectrum from deep shadow to blown-out highlight. GANs, by design, tend not to generate fully saturated or severely underexposed regions. When you check the histogram of a suspect image and find it suspiciously clipped — no true blacks, no true whites — that's a lighting signature that shouldn't be ignored.

Check 2: Source Tracing — Because Compression Is the Enemy

Here's the investigative complication nobody talks about enough: social media platforms compress images on upload. Every time a photo gets saved, re-shared, screenshotted, and reposted, it loses data. And artifact-based detection methods, tested on raw uncompressed image data, perform significantly worse on the compressed versions that are actually spreading across feeds. Research on detection performance across compressed video and images confirms this problem is substantial — compression can mask a meaningful proportion of detectable artifacts, effectively laundering the evidence.

This is precisely why artifact analysis alone is insufficient. Source tracing — examining the upload history, metadata, and origin chain of an image — becomes a parallel verification track, not a backup plan. Previously in this series: Your Face Is The New Password And Nobody Asked If It Should .

What does source tracing actually involve? At the technical level, proactive forensics approaches embed watermarks using dual-tree complex wavelet transforms in the high-frequency sub-bands of authenticated images. When that watermark is present and verifiable, you have mathematical proof of origin. When it's absent — or when the metadata trail leads to an anonymous upload with no provenance — that's evidence of a different kind. Absence of authenticated origin doesn't prove fabrication, but it removes the only mechanism that would prove authenticity.

"Visual imperfections on faces will likely disappear soon, with newer GAN architectures producing faces with even more details and highly realistic appearance, meaning relying exclusively on visual traces could be a losing strategy in the long term." — From Deepfake Media Forensics: State of the Art and Challenges Ahead, arXiv

That's the core problem stated plainly by the research community itself. The visual approach has an expiration date. Source authentication doesn't — because mathematical proof of origin is either there or it isn't, regardless of how photorealistic the face becomes.

Check 3: Cross-Image Consistency — The Pattern That Repeats

One viral image is hard to evaluate in isolation. Multiple images from the same claimed subject, or multiple suspect images from the same apparent source, tell a completely different story — and this is where investigators often find the most reliable evidence.

Here's why this works: manipulation artifacts are local in nature, meaning there's no broad semantic difference between a real image and a fake one at a glance. But those local artifacts follow consistent patterns across every image generated by the same method. When you compare a series of suspect images, the same mathematical fingerprints repeat — same frequency signature, same artifact zones, same intensity range characteristics. Research on GAN fingerprints in face image synthesis confirms that different GAN models produce distinct, identifiable fingerprints, making it possible not just to detect fabrication but to link multiple suspect images back to the same generative source.

At CaraComp, this kind of cross-image analysis — measuring pixel-space relationships using Euclidean distance across facial feature sets — is foundational to how facial comparison works forensically. The same mathematical framework that exposes deepfakes by finding inconsistencies is what validates authentic identity by finding consistencies. The methodology cuts both ways.

The Misconception That Gets People in Trouble

It's completely understandable why people trust their visual assessment of a face image. Human beings have spent their entire evolutionary history developing exactly that skill — reading faces is one of the most deeply practiced cognitive tasks we do. When a face looks right, it feels right. The lighting tracks. The skin has texture. The eyes look like they're focused on something real. Up next: Biometric Data Legislation Investigator Compliance Risk.

The problem is that GANs were built specifically to exploit this bias. The misconception isn't stupidity — it's a reasonable heuristic being applied to a situation it wasn't designed for. "Does this look real?" is the right question when assessing a photo taken by a human with a camera. It's the wrong question when the face was generated pixel-by-pixel by a model trained on millions of actual human faces.

The correct question isn't perceptual. It's forensic: Can I verify the mathematical origin of this image? And if I can't — if there's no watermark, no clean metadata trail, no provenance chain — then visual realism tells me nothing useful.

So here's the question worth sitting with: if a case came down to a single viral face image with no metadata trail and no comparison images — just one frame, visually flawless, source unknown — which of the three checks would you trust to give you a real answer? The artifact analysis that compression may have already stripped? The source trace with nothing to find? Or the uncomfortable conclusion that a perfect-looking image with no verifiable history isn't evidence of anything at all?

The most dangerous deepfake isn't the one that looks fake. It's the one that looks exactly right, uploaded anonymously, shared ten thousand times before anyone thought to ask the second question.