01 The problem
Video used to be evidence. It no longer is on its own.
For most of a century, a video of someone saying or doing something was close to proof that it happened. Generative AI has quietly removed that certainty. It is now possible to produce footage of a real person, their face, their voice and their manner, saying words they never said, at a quality that a careful human viewer cannot reliably challenge.
The task we work on is the inverse of generation: given a piece of video, decide whether a machine had a hand in making it. A synthetic clip is never a perfect copy of reality: the process that creates it leaves faint, systematic traces. Detection means finding those traces reliably, even when they are subtle, even on footage that has been compressed and re-shared, and even when the tool that made it did not exist last month.
02 Why it matters
When anything can be faked, trust itself becomes the target
The damage is no longer hypothetical. Synthetic video is already moving money out of companies, impersonating executives and public figures, and injecting doubt into elections. And there is a second-order cost: once people know video can be faked, the genuine can be dismissed as fake too. A society that cannot agree on what is real is easy to manipulate. Reliable detection is part of keeping the public record trustworthy.
Economic
Impersonation fraud scales the moment a face and a voice can be synthesised on demand. Detection is a control that protects payments, contracts and identity verification.
Democratic
A single convincing fake at the wrong moment can move an electorate. The ability to flag manipulated media quickly is now part of the integrity of public debate.
Institutional
Courts, newsrooms and administrations increasingly need to answer one question about a file: was this made by a camera or by a model? That question needs a technical, defensible answer.
03 Why it is hard
The thing you are detecting is designed to be undetectable
Deepfake detection is an adversarial problem: every improvement in generation is, by definition, an attempt to erase the traces detectors rely on. A method that works today can quietly stop working when a new generator appears, and the failure is silent: the detector keeps returning confident answers that are now wrong.
A moving target
Generators evolved from one family of techniques to another in a few years. Each generation leaves different fingerprints, so yesterday's detector is not guaranteed to recognise tomorrow's fake.
Generalising to the unseen
The generator used in a real attack is usually one the detector was never trained on. Performing well on known fakes is easy. Performing well on unknown ones is the actual research problem.
Compression erases evidence
Real footage is re-encoded every time it is uploaded and shared. That compression scrubs away exactly the fine detail a detector wants to read, so it must survive the messy quality of real-world video.
Adversarial pressure
An attacker who knows a detector exists can nudge a fake specifically to slip past it. Robustness has to hold up against someone actively trying to defeat it, not just against average content.
The cost of being wrong
Calling real footage fake is as harmful as missing a fake. A usable detector has to be honest about its own uncertainty, not just output a verdict.
Speed at scale
The volume of media that needs checking is enormous and arrives fast. Detection that only works in a lab, slowly, does not help the newsroom or the fraud desk that needs an answer now.
04 Our angle
Detect the traces, and hold up as the fakes improve
We work on detecting AI-generated video by the traces that generation leaves behind, with the explicit goal of staying robust as generators get better. The interesting problem is not beating today's fakes, but building detection that does not quietly collapse the moment a new generator appears.
This line is early and stated as such: no detection performance is claimed here, and none will be until the approach survives evaluation on data hard enough to mean something. The approach itself is not described at this stage.