THE POWER GRID'S HUM, A WITNESS AGAINST AI VIDEOS

Detecting AI-generated video has become an arms race, and according to Ryan Wyton and Bacha Rehman of Southampton Solent University, it is one “the detector is losing”. Most detectors learn the telltale fingerprints of the generators they were trained on. On a recent benchmark of six million clips, the best reach only around 80% accuracy overall, and fall to near or below chance on generators they have never seen.

The two researchers looked instead for a signal that generators do not control: the electric network frequency, or ENF.

The grid’s fingerprint

Mains electricity is nominally 50 or 60 hertz, but its exact frequency drifts by small amounts as supply and demand change. Because a power grid is tightly coupled, this drift is the same, to within milliseconds, across the whole network at any instant — so it acts like a timestamp. It gets into recordings in two ways. Microphones pick it up as a hum. And lamps powered by the mains flicker at twice the grid frequency, 100 or 120 hertz. A camera whose sensor reads the image line by line (a “rolling shutter”) records that flicker as faint bands in the video.

A premise that did not survive

The work began with a simple idea: synthetic video would have no grid flicker and could be flagged by its absence. It “did not survive contact with the evidence”, the authors write. Across 600 clips from six text-to-video generators, the flicker appeared as strongly as in real footage — with an apparent frequency set by the generator’s frame rate, not by any real grid. Measured on flicker strength alone, real and generated clips could not be told apart: a score of 0.51, where 0.5 means pure guessing.

So the question changed. Instead of asking whether a video has a grid signal, the system asks whether that signal matches an independent record of the grid at the claimed time and place. A matching trace is strong evidence of a genuine recording. A missing one proves nothing, since many real recordings — outdoors, dim, or made with professional sound gear — carry no usable hum.

How the system works

One small neural network (about 157,600 parameters) reads the grid signal in both sound and video.
Real labelled examples are scarce, so the team trained it on synthetic ones: they modelled how grid frequency wanders and is pulled back towards its nominal value, then injected that drift into hum-free speech and flicker-free video, along with five kinds of inconsistent signal (absent, flat, broken, wrong grid, spliced).
A separate step matches the recording’s signal against a reference log of the grid, either to check a claimed time or to find an unknown one.
A fixed decision graph — the “agentic” part — combines this with existing detectors of manipulated faces, lip-sync and metadata, calling the slow ones only when the quick checks are inconclusive.

The results

On real recordings from sources never seen in training, the system accepted 93% of genuine audio and 100% of genuine videos, against 53% and 15% for a naive “is there a hum?” test.
Checking a claimed recording time separated genuine from false timestamps perfectly on the recordings where a trace could be extracted; a blind search found an unknown recording time to within about one second in a 24-hour reference.
In a real scene, the video’s flicker matched its grid reference and the same recording’s sound. Two AI recreations of that scene flickered at 60 hertz, the wrong grid for a location claimed to run at 50.
Among 91 genuine recordings checked against their true grid, it raised no false accusation.
Running the slow detectors only when needed cut computing time by 54% without losing accuracy.

The hard case

The grid signal authenticates the recording, not what it shows. If someone films a real scene and then swaps a face into it, the hum stays genuine. That falls to the face and lip-sync detectors, which on compressed footage of real face-swaps performed close to chance: the full ensemble scored around 50%. The method also needs grid reference logs, which the authors describe as public for the UK and continental Europe but restricted or unpublished elsewhere, and it cannot stop a forger who splices in a genuine grid recording from the right time and place.

The authors pitch their system as a triage tool, one line of evidence among others. Its goal, they write, “is not to declare a clip fake, but to say, with calibrated confidence, when a recording is what it claims to be.”