Deepfake Detector

A deepfake detector is a useful first filter, not a verdict. The strongest check combines a detector score, a close look at the frames and the oldest copy of the file you can trace.

01How a deepfake detector reaches its score

Most detectors are classifiers trained on large sets of real and manipulated media. They look for traces a generator leaves behind: blending seams around the face, skin and hair texture that looks too smooth, compression that differs between the face and the background, or audio patterns typical of voice cloning. The result is a probability, often shown as a percentage and sometimes with a heat map of suspicious areas.

That score answers a narrow question: does this file resemble the fakes the model learned from? A new generator can slip past an older detector, and a heavily compressed WhatsApp forward confuses many tools because compression erases the fine traces they rely on. Provenance checks, covered below, work differently: they read a signed record or a hidden watermark instead of judging how the pixels look.

02How accurate are deepfake detectors

Less accurate outside the lab than headline numbers suggest, because real fakes rarely resemble the test sets detectors are scored on. Two public benchmarks show the gap.

The same 2025 study found that commercial detectors and models retrained on recent fakes did better but still trailed trained forensic analysts. In practice: run more than one kind of check, and never post a detector percentage as proof that a clip is fake.

03Free deepfake detectors worth trying

No free tool is reliable alone. Combine a classifier, a provenance check and a reverse image search, since each catches problems the others miss.

• Hive AI Detector: a free Chrome extension from Hive that scores images, video, text and audio, and guesses which generator made an image or clip.
• SynthID check in the Gemini app: upload one file and ask whether Google AI made it. It is free for signed-in users, takes files up to 100 MB and videos under 90 seconds, and only recognises Google's own watermark.
• InVID-WeVerify plugin: a browser toolbox maintained by AFP Medialab that splits videos into keyframes for reverse image search and reads image metadata. A few advanced tools are reserved for journalists and researchers.
• Content Credentials verifiers: read the C2PA record that some cameras, editing apps and AI generators attach, showing how a file was made and edited. A missing record proves nothing, since many platforms strip metadata on upload.

Every upload to a checker creates a copy on someone else's server, so read the tool's privacy terms before submitting a clip of a real person. The questions in the face swap safety guide apply here too: how long the file is kept and whether it trains the model.

04How to spot a face swap in a video

Video face swap tools rebuild the face frame by frame, so the seams show up in motion rather than in a single still. Watch at full resolution, slow playback to 0.25x in YouTube or your phone's video editor, and pause on these moments:

• Head turns: the face slips, blurs or changes shape as the person moves into profile.
• Mouth and teeth: lips drift out of sync with speech, and teeth merge into a pale block.
• Edges: the jawline, ears and hairline flicker, double up or smear against the background.
• Objects in front of the face: a hand, microphone or pair of glasses warps the features as it passes.
• Light and skin: tone and shadows on the face do not match the neck, hands or room.
• Eyes: blinking looks rare or mechanical, or reflections differ between the two eyes.

One odd frame proves little, because compression and beauty filters create similar artefacts. Several signs together, plus a source you cannot trace, are a strong reason to doubt a clip. Short loops are harder to judge, since GIF face swaps have few frames and heavy compression, and in group swaps each face is worth checking, because one replaced face may carry different lighting from the rest. On a live video call, ask the person to turn fully sideways or pass a hand across their face; real-time swaps often break under both.

A detector score is a lead to follow, not a conclusion to share.

05What to do if you find a deepfake of yourself

Platforms, police and regulators in Malaysia ask for the same basic evidence, and a complete report moves faster.

1. Save the evidence: the link, account name, date and time, plus screenshots or a screen recording of the post and its comments.
2. Report the post to the platform as impersonation or manipulated media. Since 1 January 2026, the Online Safety Act 2025 has required licensed platforms to offer tools for reporting harmful content.
3. If anyone asked for or lost money because of the fake, call the National Scam Response Centre on 997 immediately, since early reports give the best chance of freezing transfers.
4. Lodge a police report and a complaint with MCMC, attaching the saved evidence.
5. Warn family, friends and colleagues so they ignore calls or messages that appear to come from you.

Do not message the uploader or reshare the fake to debunk it, as that spreads the clip further. If the content harms your reputation or business, a lawyer can advise on a defamation claim. Many face swap apps need only one clear photo, so if you make swaps yourself, use only your own face or faces of people who agreed to it.

06Deepfake detector FAQ