A conventional approach to deepfake detection is relatively straightforward: show a machine-learning model thousands or millions of examples labeled “real” and “fake” and let it learn the patterns that separate them. 

The problem is that the model decides which patterns matter — without telling you why.

Imagine training a system to identify your daughter from thousands of photos, but every training photo of her happens to be taken in your apartment. The model may seem highly accurate, but you don’t know whether it learned her face or simply learned the background, lighting, and furniture that appears in all of those photos.

Take your daughter outside and the system may suddenly struggle to recognize her. Put someone else in the same apartment and it may make the opposite mistake, identifying them as your daughter.

The system may have appeared to work in testing, but for the wrong reason.

The same problem can occur in deepfake detection. A model may learn the fingerprint of a particular generator, characteristics of its training data, compression artifacts, or other correlations it guesses will distinguish the “real” and “fake” examples it has seen. It may perform extremely well on a benchmark test without learning something fundamental about how the manipulation was actually created. And it can’t tell you why it made the guess it did — that’s why we call it a black box.

At GetReal, we start somewhere different: with the evidence. Rather than asking whether a video simply “looks real” or “looks fake,” my team of researchers asks: What has this content been through and how has it been transformed?

Every digital image, video, or audio stream has a history. From the moment it’s created by an AI model or recorded by a camera or microphone, it passes through a series of mathematical transformations before reaching the person on the other side of a call. What you see on the other side of a video call is not a person, but pixels that have been mathematically transformed and reassembled. 

Compression, resampling, denoising, re-encoding, and manipulation all leave traces, and we study these traces. We’re able to tell the model exactly what to look for. 

We start by studying the adversary: how face swaps, voice clones, avatars, and other AI-powered impersonation systems actually work, researching the underlying mathematical operations they share.

Consider a real-time face swap. Different tools may produce very different results, but they face the same fundamental problem: continuously transforming one person’s face to follow another person’s head pose, expressions, and movements. That requires underlying processes such as image warping, blending, or resampling.

Instead of asking whether a video resembles examples produced by a specific tool, we look for evidence created by the underlying process itself.

Next, we study the environment in which the attack has to operate.

A live video call is a highly dynamic environment. Cameras, microphones, devices, lighting, bandwidth, and conferencing platforms all transform the underlying media. Zoom, Teams, and Webex process streams differently. Bandwidth fluctuates. Frames drop. Audio is resampled and denoised.

Those transformations also leave traces — but they aren’t attacks. And you cannot reliably identify what is abnormal until you understand what normal looks like.

By looking for traces and identifying which ones we want to look for, we don’t have to retrain the models every week or every month. Our researchers therefore model how authentic streams behave across real-world communications environments. That gives us the context to distinguish normal variation from evidence of manipulation.

This is also what makes the GetReal approach resilient as deepfake technology evolves. When a new face-swapping tool appears, the question isn’t simply whether we’ve trained a model on examples from that particular tool. The question is whether the new technology still leaves the underlying forensic traces our research has already identified.

And when something genuinely new appears, our Threat Intelligence and Research teams can examine how it works, identify what changed, and extend our detection accordingly.

This is why we start with the evidence and not the verdict. When you understand what a piece of content has been through, and compare it to what normal looks like, manipulation has no place to hide. That’s not a trick that breaks down, but a resilient, sustainable method that holds regardless of what comes next.