How to Detect Deepfakes: Deepfake Detection Beyond Eye Tricks

Here's something that should stop you mid-scroll: the production team behind the Peacock series Ted tried every normal trick to turn Seth MacFarlane into Bill Clinton. Makeup. Prosthetics. Traditional CGI. All of it failed, not because the artists weren't good, but because MacFarlane's skull and Clinton's skull are shaped differently enough that no amount of surface-level work could bridge the gap. The result, every time, was what VFX professionals call "the uncanny valley", close enough to human that your brain screams wrong before you can even say why.

So they built a deepfake instead. And it worked. Audiences watched it and believed it.

That moment, audiences fooled by a synthetic face in a major TV production, is a useful window into something most people haven't figured out yet. The old advice about spotting deepfakes ("watch for blurry mouths, weird blinking, glitchy edges") was written for a different era. The technology has quietly outgrown those tells. And the reason it has isn't magic. It's layers.

---

Why Your Brain Gets Fooled: The Surface vs. the Structure

Detection Methods That Watch Audio, Not Just Faces

Before we go further, it helps to define detection methods in plain terms. Detection methods are the tools and techniques used to identify deepfakes, and the strongest ones don't just watch a face, they watch audio too. A voice track that doesn't quite sync with mouth movement, or audio that sounds slightly flat compared to the room it was supposedly recorded in, can be one of the clearest deepfake detection signals available.

Most people think of a deepfake the way you'd think of a Snapchat filter, a digital mask dropped over a face in real time. Swap the face, done. Easy.

That model is about five years out of date.

The real process, especially for anything convincing enough to fool a general audience, breaks into at least two completely separate workflows. Think of them as two film crews shooting the same scene. One crew is capturing motion. The other is capturing appearance. Neither crew's footage is the finished product. The magic (and the deception) happens when someone fuses the two together, frame by frame, at 24 to 30 frames per second.

Here's what that actually looks like in practice. This article is part of a series, start with How Deepfake Video Detection Actually Works.

For the Ted/Clinton sequence, the production team used a six-camera rig pointed at MacFarlane's face while he performed. Six cameras, not one, not two, because replacing a face computationally requires a level of geometric detail that a standard film camera simply can't capture on its own. You need to see the face from enough angles simultaneously to reconstruct its three-dimensional shape with precision. That data, thousands of points mapping exactly how MacFarlane's face moved, stretched, and wrinkled during each take, became the foundation for everything that followed.

That's Pipeline One: motion capture. It records what the face does.

Pipeline Two is face replacement, taking a 3D scan of the target face (Clinton's, in this case, built from photographs and archival footage) and computationally fitting it over the motion data from Pipeline One. The replacement face has to match the lighting of the scene, the texture of skin under that specific light, and, this is the part that breaks most attempts, the edges.

---

The Edge Problem Nobody Talks About

Edges are where deepfakes die.

When you replace a face frame by frame, you're constantly managing the boundary between the synthetic face and the real everything-else, the neck, the ears, the hair, the background. If the original actor's mouth is open and the replacement face has a slightly different mouth shape, you get a ghost: a faint outline of the original that bleeds through. Fix it on one frame, and the problem moves to the next one.

The technical fix is called inpaintingbasically, teaching the software to intelligently fill in what should be there based on surrounding visual information, the way a photo editor might clone-stamp a background to cover something up. But inpainting has to work across every single frame in a sequence, and it has to do it without creating a flicker or a "swimming" artifact (that subtle rippling effect that older deepfakes always seemed to have around the jaw).

And that's still not the hardest part.

The hardest part is temporal coherencekeeping the face consistent not just in one frame, but across hundreds of consecutive frames in motion. One perfect-looking still image is achievable. Maintaining that same quality at 24 frames per second, through head turns and expressions and lighting changes, is an entirely different problem. It's one reason why, according to Beverly Boy Productions, a fully unified face replacement pipeline has to solve for geometric alignment, surface texture, lighting, and expressive realism all at once, simultaneously, continuously, across the entire sequence. Previously in this series: Your Boss Just Called You On Video It Wasnt Him 25m Is Gone.

"We exhausted all practical methods, makeup, prosthetics, before turning to AI-driven face replacement." Seth MacFarlane, as reported by The Hollywood Reporter

Let that sink in. A professional production, with a budget reportedly ranging between $8 million and $10 million per episodecouldn't pull off a convincing face replacement through traditional means. They needed AI-driven compositing to close the gap. And when they did it right, nobody in the audience could tell.

---

The Misconception That's Getting People Fooled

Detect Deepfakes By Watching Structure, Not Surface

If you want to detect deepfakes reliably, you have to stop scanning surfaces and start thinking about structure. Media forensics teams that specialize in deepfake detection rarely rely on a single frame or a single clue. Instead they use detection techniques that combine facial geometry, motion consistency, and audio analysis into one judgment, because any one signal alone can be faked.

Here's where people get tripped up, and honestly, it's not their fault.

The "red flags" list that circulated a few years ago, look for blurry edges around the face, unnatural blinking, weird skin texture, mismatched lighting, was genuinely useful when those things were genuine weaknesses. Early-generation deepfakes failed at surfaces. So people learned to scan surfaces.

But the real challenge in synthetic video was never the surface. It was always the structure underneath: the geometric relationship between facial features, the way those features move through time, the consistency of how light hits a moving face across dozens of frames per second. Nobody taught the public to look for those things, because those things are invisible without specialized tools.

Think of it like this. Imagine someone restoring a damaged oil painting. A beginner restorer might dab matching paint over the damaged area and call it done, and from across the room, it looks fine. But a conservator (a museum painting expert) knows to look at the underlying brushstrokes, the way paint layers built up over time, the micro-texture of the canvas showing through. The surface can be perfect. The structure underneath tells the real story.

Modern synthetic video is the same. The surface, skin texture, eyelash detail, lip movement, can be rendered convincingly enough to pass a casual glance or even a careful one. What gives it away, when anything does, is structural: a subtle wrongness in how the face responds to moving light, a fractional inconsistency in how the jawline tracks with head movement. These are things your eye picks up as a vague feeling of off without being able to name why.

That vague feeling is not a reliable detection method. Up next: That Urgent Video From Your Boss Your Eyes Cant Catch The Fa.

---

So What Do You Actually Do With This?

How Deepfake Detectors Actually Work

A deepfake detector algorithm can be trained on thousands of real and synthetic clips until it learns common cues that separate the two, patterns humans simply can't see by eye. Some deepfake detectors run temporal and spatial analysis together, checking both a single frame and how that frame changes across the next several frames. Even so, deepfake detection is opaque to most people, which is exactly why understanding the underlying media and technology matters more than trusting a gut reaction.

The practical takeaway here isn't "learn to spot deepfakes visually", because that bar keeps moving, and it's moving faster than any checklist can keep up with. The production team behind Ted Season 2 said explicitly that for synthetic video to succeed, it has to look absolutely real, and they built an entire specialized pipeline to achieve exactly that.

If teams with that kind of budget and technical depth are building synthetic video that's designed to be undetectable by eye, then the safest assumption when a video matters, when it's making a claim, asking for money, showing someone doing something they'd deny, is that looking convincing is not the same as being real.

At CaraComp, this is exactly the gap we work in. Facial recognition and identity verification aren't just about matching faces. They're about understanding the structural and temporal signals that your eye can't consciously process, the kind of layered analysis that matches how synthetic video is actually built.

For everyone else, the parent, the employee, the person who just got forwarded a wild video at 11pm, the rule is simpler: slow down, check the source, and treat visual evidence the same way you'd treat an unsigned check. It might be real. But "it looks real" is not the verification. The source is the verification.

Here's the question worth sitting with: If a six-camera rig and a team of specialized artists spending millions of dollars per episode couldn't make a convincing synthetic face with traditional methods, but did succeed with AI-driven layered compositing, what does that tell you about videos shot on a phone, with tools that cost nothing, by someone with a reason to deceive you?

The Ted production was trying to entertain you. Not everyone building synthetic video is.

Learning how to detect deepfakes means accepting that no single check is enough on its own. Deepfake detection works best as a layered process, much like the layered pipelines used to build the fakes in the first place. Detection signals from audio, motion, and lighting all get weighed together rather than treated as one make-or-break test.

If you are trying to spot deepfakes in something sent to you directly, start by asking where else that same video or claim appears. A deepfake detector, whether human or software, works better when it has more than one piece of media to compare against, since deepfake detection improves a lot when there's context around the clip and not just the clip itself.

Detection in professional settings usually combines several detection methods rather than picking one favorite tool. Forensic teams look at unnatural motions in the jaw and neck, check whether audio timing lines up with lip movement, and run the footage through models built specifically to flag detection signals a person would miss. This layered approach is slower than a glance, but it is far more reliable than trusting how convincing a video feels.

None of this requires special equipment for an everyday viewer. The technology behind modern detection methods is complex, but the habit it points to is simple: treat a single viewing as the start of the process, not the end of it, and let media forensics tools do the structural checking your eyes physically cannot do.

Detect Deepfakes With Layered Detection Signals

Anyone trying to detect deepfakes with layered detection signals needs to look past the picture and into the technology that built it. The most useful checks combine facial geometry, motion tracking, and audio timing, because relying on any single clue leaves an easy blind spot. Modern technology makes it possible to weigh several of these signals at once instead of judging a clip on looks alone.

Deepfake Detection Also Relies on Audio

Deepfake detection also relies on audio because a mismatched voice track can expose a fake long before the visuals do. The technology used to analyze sound looks for flat room tone, timing gaps between lips and words, and other clues that a microphone did not actually capture that voice in that space. Pairing audio checks with visual checks gives a fuller picture than either one alone.

Why Deepfake Detection Still Matters

Deepfake detection still matters because the underlying technology keeps improving on both sides, the side making fakes and the side catching them. Researchers build models that study thousands of clips to learn which patterns separate real footage from synthetic footage, and those models get retrained as new techniques appear. Staying current with the technology is less about memorizing a checklist and more about trusting a layered process.

None of this technology works in isolation. A single tool might flag a lighting inconsistency, while another checks how sound lines up with lip movement, and a third looks at how a face moves across many frames of media. Combining outputs from different tools is what turns raw technology into a dependable answer, one that a casual glance simply cannot match.