Deepfake Detection: Why Experts Assume Video Is Fake

Deepfake detection and video provenance concept

Hany Farid, a computer science professor at Dartmouth College who has spent three decades building deepfake detection techniques, now says people should treat video as fake until it is shown to be real. He calls this a presumption of fake, and he argues it points toward a future where most online media can be trusted again.

For most of his career, Farid says, he and nearly everyone else assumed recordings were authentic. If a skeptic called a video fake, the skeptic had to prove it. Generative AI has reversed that burden. Now the creator has to prove the content is real.

The Tomahawk video that changed his mind

The shift came on a Saturday in March, when reporters from around the world contacted Farid. An Iranian news agency had aired a week-old video that appeared to show a US-made Tomahawk missile hitting a school. Reports said the strike killed more than 175 people, most of them children. The US president and military denied responsibility.

Farid was suspicious. The quality was low, the footage emerged weeks after the bombing and a single agency with a weak record for accuracy was distributing it. Still, the stakes were high, so he analyzed it.

He aligned the frames against a fixed landmark to remove camera motion. If the missile had been pasted in without regard for that motion, its path would look physically wrong. Instead it moved in a nearly straight line. He then estimated the missile’s size in pixels and found it matched a 6-meter Tomahawk at roughly 100 meters. A 0.3-second gap between the visible impact and the sound fit that distance, given sound travels at 343 meters per second.

After several more hours, he judged the video authentic. The experience taught him something about himself. A recent study he worked on with Berkeley colleagues showed that people are now so wary of video that they struggle to believe even real clips. Farid had behaved like the study participants. A lack of visible artifacts no longer convinced him.

Provenance signals

Forensic analysis like his does not work at internet speed. Provenance signals do. They include content credentials and watermarks added when media is created or recorded.

Credentials built on the C2PA standard attach a cryptographically signed manifest to a photo or video at capture. It records the device, the time and location and any later edits, and machines can detect tampering with either the content or its history. Watermarks such as Google’s SynthID embed an imperceptible signal in pixels or audio samples that identifies where content came from.

Apple has taken the idea further. On some of its newest iPhones, the camera sensor can sign raw pixel data as it is captured. That data is developed into a reference image stored beside the normal, editable photo, and anyone can compare the two to see what changed.

Critics call these signals partial and open to attack. Farid agrees if they are viewed as a way to catch bad actors, but he says that isn’t their job. They let honest creators prove their recordings are genuine. The goal is to move from catching bad actors to backing good ones, and if enough good actors adopt them, they can drown out the bad.

Fake video calls

The same distrust now reaches ordinary video calls. In September 2024, Senator Ben Cardin, then chair of the Foreign Relations Committee, got a meeting request from Dmytro Kuleba, who had been Ukraine’s foreign minister until a few weeks earlier. Cardin grew suspicious when the caller pressed him on whether the US should let Ukraine fire long-range missiles into Russia. He ended the call. The State Department later confirmed that Kuleba never asked for the meeting, and the Senate security office warned staff to expect similar attempts, possibly from Russia.

Such impersonation needs only a face-swap or avatar deepfake plus voice conversion, both available in commercial products and open-source libraries and able to run on a modern laptop. Real-time fakes face limits, though. They must render at around 24 frames per second, so software takes shortcuts, such as simplifying faces and allowing drift between audio and video. Eyes miss these, but forensic tools can expose them.

Where verification goes next

Farid sees a role for signing in live calls. Video software could sign each frame of a stream as Apple now signs stills. A face-swap happens between camera and conferencing app, so the software could check whether incoming frames were signed by a physical sensor or altered in transit.

His conclusion is that forensics was enough for the first two decades of the digital era, but AI changed the scale and speed of manipulation. To be believed, a recording now needs a provenance signal from the moment of capture. Provenance lets good actors be trusted, and forensics roots out the bad ones.

Source: IEEE Spectrum

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.