Deepfake detection: Can automated software prove if a video is synthetic?

A sensational video starts circulating on social media showing a prominent government official making a shocking confession, or a company executive announcing an improbable corporate merger. Within minutes, comment sections erupt into fierce arguments: half the viewers claim the video is undeniable proof of wrongdoing, while the other half insist it is a computer-made deepfake. Someone downloads the clip, uploads it to a free online scanner, and posts a screenshot of the result: '94% Probability of Synthetic Manipulation.'

The idea of a single automated tool that can instantly tell truth from fiction sounds wonderful. But extensive testing by cybersecurity researchers and institutions like DARPA shows that automated deepfake detection is an ongoing technical game of cat-and-mouse (an adversarial arms race), not an infallible digital lie detector.

The forensic signals detectors analyze

Automated forensic tools inspect suspicious video files across several physical and mathematical layers:

• Natural bodily signs (physiological signals): Living human bodies make tiny, involuntary movements every second. With every heartbeat, blood flows through tiny facial vessels, creating microscopic shifts in skin tone that cameras can detect (a technique called remote photoplethysmography). Early faked videos often failed to simulate these subtle pulse rhythms or natural blinking habits.
• Mismatched lighting and reflections: In authentic video, tiny glints of room light bouncing off the glossy surface of both eyes (corneal specular reflections) match the surroundings perfectly. Face-swapping programs often paste a face recorded under bright studio lamps onto a background filmed in afternoon sun, creating clashing shadow angles.
• Edge wobbles between frames (temporal boundary jitter): When a person moves or turns their head, detectors track fine edges like teeth, earlobes, and eyeglass frames. Artificial intelligence models frequently introduce slight flickering or blurring along the outer borders of the face from one split-second to the next.
• Invisible mathematical fingerprints (frequency domain artifacts): Image-generating software builds pictures by resizing mathematical grids, leaving behind subtle, repeating patterns hidden deep within the image data (in mathematical frequency space) that specialized forensic classifiers can spot.

Acoustic detection and the voice challenge

Verifying audio recordings comes with just as many hurdles. While audio detection programs listen for missing breaths, unnatural pitch jumps, and sound mismatches between stereo channels, modern voice synthesis software (diffusion vocoders) sounds remarkably natural. In addition, when phone calls are compressed over cellular networks, subtle background frequencies get stripped away to save bandwidth. This normal compression erases the very acoustic clues that detectors rely on, making audio forensics especially challenging on voicemails and phone calls.

The compression bottleneck on social media

Under clean laboratory conditions with crystal-clear video, forensic classifiers score high marks. But out in the real world, detectors hit a massive obstacle: social media video compression.

Whenever a video is posted to apps like WhatsApp, TikTok, or social networks, servers compress the file heavily to save bandwidth. Video-shrinking formulas (compression codecs) lower resolution, soften textures, and remove fine color shifts. This process accidentally erases the faint skin pulse signals and subtle mathematical patterns that detectors need, causing tools to flag real videos as fake (false positives) or miss actual deepfakes entirely (false negatives).

The perpetual generator arms race and provenance solutions

Furthermore, detection tools are locked in an endless cycle called an adversarial arms race. The moment researchers publish a new way to catch deepfakes—like spotting mismatched reflections in someone's eyes—AI creators feed that exact test into the model's training rules (its loss functions). The software quickly learns to fix that flaw, eliminating the telltale sign.

Because automated scanning tools can never be completely foolproof on heavily shared video files, cybersecurity experts are turning toward digital certificates of origin (cryptographic provenance). Industry standards like C2PA Content Credentials embed tamper-evident digital signatures directly into files at the camera level, documenting when and where a clip was recorded. Until this technology becomes widespread, the most reliable way to handle shocking online videos is to check multiple independent news sources and established reporting, rather than trusting a single automated scanner score.