What is a deepfake and how to verify it step by step

1. Face swap. Replacing a person's face with another in a video. The oldest technique — popularised by DeepFaceLab and FaceSwap.

2. Face reenactment (puppetry). Keeping the target's face but driving their expressions and mouth movements from a source actor. Used in most political deepfakes.

3. Lip-sync. Modifying only the mouth region so a real video appears to say words the subject never uttered. Extremely hard to spot.

4. Full-body synthesis and voice clones. Text-to-video (Sora, Runway Gen-4) and voice cloning (ElevenLabs, OpenVoice) now generate entire scenes from a prompt — the hardest to detect and the fastest-growing category in 2026.

In February 2024 a Hong Kong finance employee transferred $25 million after a video call with a deepfake CFO. In March 2024 a robocall cloning Joe Biden's voice asked New Hampshire voters not to vote in the primary. In 2025 deepfake-assisted fraud losses reached an estimated $40B worldwide according to Deloitte. The 2024 EBU report found 77% of European broadcasters mistakenly aired AI-generated content.

Blink patterns. Early deepfakes blinked too little; modern ones sometimes blink too mechanically.

Temporal flicker. Face swaps leak at the edges: hairline, ears, neck.

Lighting mismatch. The pasted face rarely matches the scene's light direction and colour temperature.

Pupil reflections. Real pupils reflect the same environment; deepfake pupils often show different reflections in each eye.

Audio-visual sync. Lip-sync deepfakes drift by 30–80 ms on long sentences.

Step 1 — EXIF and container analysis. Check encoding software, timestamps, container structure.

Step 2 — Forensic tool. Run it through ScanTrace or equivalent. Record the verdict and confidence.

Step 3 — Reverse search. Google Lens, TinEye, Yandex. Find the oldest occurrence online.

Step 4 — Source contact. Verify with the person or organisation allegedly in the media. A 30-second phone call beats any detector.

Step 5 — Cross-reference. Check official social accounts, press releases, agenda of the person. If the statement contradicts known positions, treat as suspicious until proven otherwise.

ScanTrace — image + video, Spanish/English UI, PDF certificate, free tier. Reality Defender — enterprise-grade, real-time. Intel FakeCatcher — real-time for live streams. Sensity AI — deep catalogue of known deepfake campaigns. DeepFake-o-meter (UB) — open-source research tool.

The single biggest mistake in deepfake verification is trusting any tool's verdict in isolation. Tools fail; protocols don't — if they're followed. The five-step protocol above, combined with ScanTrace for the forensic layer, brings false-publish rates below 1%. Try the free deepfake checker to run step 2.

What exactly is a deepfake?

A deepfake is synthetic media — image, video or audio — where a real person's face, voice or body has been replaced or generated by a deep learning model. The term combines 'deep learning' and 'fake'.

Is every AI-generated image a deepfake?

No. A deepfake strictly refers to content that impersonates a real, identifiable person. A landscape generated with Midjourney is an AI image but not a deepfake.

How accurate are deepfake detectors in 2026?

Top forensic detectors reach 92–97% accuracy on current models. That remaining 3–8% is exactly why any editorial decision still needs a human in the loop plus multiple independent signals.

Can deepfakes be detected in live video calls?

Partially. Tools like Reality Defender and ScanTrace's video endpoint flag temporal inconsistencies — flickering, mismatched blinking, unnatural head turns. Real-time detection is still imperfect against top-tier face-swap models.

What is the legal status of deepfakes?

The EU AI Act (2024) mandates clear labelling of AI-generated content. In the US, several states criminalise non-consensual sexual deepfakes. In 2026 most democracies treat malicious deepfakes as fraud, defamation or election interference.