Deepfake detection
Deepfake detection is the use of AI and forensic techniques to identify videos, images or audio that have been synthetically generated or manipulated, such as swapped faces or cloned voices. Detectors look for clues like lighting mismatches, unnatural facial movement, compression artifacts and audio inconsistencies. Some systems also check provenance data, such as content credentials, that records how a file was created.
Why it matters
A single fake video can damage a brand, sway public opinion or enable fraud, such as a cloned executive approving a payment. Publishers risk their credibility if they air manipulated footage, and identity checks fail if fakes pass. Detection and provenance checks reduce those risks.
How it works
- Visual models scan frames for signs of manipulation, such as odd blinking, blurred edges or lighting mismatches.
- Audio models check for traces of voice cloning and for speech that does not match lip movement.
- Provenance tools read content credentials, if present, to see how the file was made and edited.
- Signals are combined into a risk score, and high-risk content goes to human verification.
Can deepfakes be detected?
Often, but not every time. Detectors catch many fakes by spotting visual and audio artifacts, but each new generation tool hides those artifacts better, and compression can fool detectors. That is why many organizations pair detection with provenance, which records where a file came from, and with human verification for high-stakes footage.
How can you tell if a video is a deepfake?
Look for faces that blur at the edges, lighting that does not match the scene, unnatural blinking or teeth, and lips that drift out of sync with speech. Check the source: who posted it first and whether trusted outlets carry it. For important cases, use a detection tool and look for content credentials that record the file's origin.
Where it's used
News publishers use deepfake detection to verify viewer-submitted footage before reporting, especially around elections. Creator platforms flag deceptive synthetic media, and identity verification apps confirm a video selfie shows a real person.
By the numbers
- The term deepfake originated in late 2017, from a Reddit community where users shared face-swapped videos. (Deepfake (Wikipedia))
- The Coalition for Content Provenance and Authenticity (C2PA) was established on February 22, 2021, with Adobe, Arm, BBC, Intel, Microsoft and Truepic as founding members. (Content Authenticity Initiative (Wikipedia))
