How Deepfake Detection Works: Detecting AI-Generated Media

How deepfake detection works

Deepfake detection does not look for the obvious. Rather than hunting for visible flaws, which modern synthetic media rarely shows, detection models look for the statistical traces that generative models leave behind: artefacts that are invisible to people but consistent across synthetic media. In short, it is AI detecting AI.

Those traces exist because a generative model, however good, reproduces the patterns of real media imperfectly. Detection models are trained to recognise exactly those imperfections.

A multi-model approach across modalities

Different media types carry different traces, so detection uses different models for each, then combines them.

Audio. Models learn from spectrogram-like representations of sound to detect manipulation artefacts across frequency ranges. Synthetic speech carries anomalies at low and high frequencies that are inaudible to the human ear. A strong approach is two-stage: the model first learns what natural speech looks like, the relationship between how someone speaks and what they say, then flags when that relationship breaks. Because it learns the shape of authentic speech rather than memorising known fakes, it can catch attack methods it never saw in training.

Image. Models analyse pixel-level noise, texture and spatial relationships across the whole image. Diffusion models, GANs and face swaps leave structural inconsistencies even when the image looks flawless to a person, and those inconsistencies are what the models score.

Video. Video detection extends image analysis across time. On top of the per-frame signals, it looks for temporal instability, flicker and frame-to-frame inconsistencies that a genuine recording would not contain.

How the models stay current

Detection only holds up if it keeps pace with generation. Three things keep it current:

• Diverse training. Models are trained on a wide range of real and manipulated media, including real-world distortion such as compression, noise, lighting and codecs, so they hold up on messy media in the wild, not just clean laboratory samples.
• Ensemble fusion. Multiple models are combined so that no single detector is a single point of failure. The ensemble is more reliable than any one model alone.
• Continuous retraining. As new generation techniques appear, detectors are retrained against them.

Crucially, effective detection is model-agnostic. It detects AI-generated content regardless of which generator produced it, and it does not depend on provenance metadata, C2PA or watermarks. It reads the media itself.

How results are delivered

Detection does not return a binary pass or fail, because the honest answer is rarely absolute. It returns a probability-based risk score: an ensemble manipulation probability, mapped to authenticity bands and a 0 to 100 severity score, with a per-signal breakdown showing which models drove the result.

This is how Signal by AI Uniti surfaces authenticity. Media collected in brand and executive monitoring is scanned, and suspicious media can be scanned on demand through Defend, part of the Unite defence layer. In both cases the authenticity verdict sits alongside the behavioural and coordination signals for the campaign around the media, not as an isolated score.

Detection plus coordination

Authenticity of the media answers one question: is this real. Behavioural detection answers the other: who is pushing it, and are they acting together. A manipulated clip amplified by a coordinated inauthentic network is a different, larger threat than the same clip sitting still, and the response is different too.

That is why content authenticity and behavioural detection work best on one platform. To understand what a deepfake is in the first place, see what is a deepfake. To see detection applied to protecting named leaders, see executive protection.