The tells that still give a deepfake away
Watch the edges and the motion: hair against a background, the line of the jaw, glasses and earrings that flicker, teeth that blur when the mouth moves. Lip movement that drifts out of sync with the words, blinking that is too regular or too rare, and skin that stays oddly smooth under changing light are common signs. On the audio side, a flat emotional tone, odd breathing and a voice that never quite matches the room are worth pausing on.
None of this is conclusive. A tell confirms a fake when it is there, but a clean clip does not prove anything, because the better tools fix exactly these artefacts. Treat the checklist as a quick screen, not a verdict.
Why your eyes are not enough anymore
Each generation of models removes the last batch of giveaways, so the gap between a real clip and a good fake keeps closing, especially after a video is compressed by a messaging app or a social platform, which hides the very artefacts you were looking for. This is where an AI media detector helps: it scores a video frame by frame for signs of generation and returns a probability, which is useful for triage.
A probability is not proof, though. A high or low score should inform a decision, not settle it on its own, so pair it with context like who sent the clip, why, and whether the story around it holds up.
Audio is the other half of the fake
A convincing deepfake is rarely just the picture. The voice carries as much of the deception, and cloned speech has caught up fast: a few seconds of a real recording is enough to generate new sentences in the same voice. On a phone call or a voice note, there is no face to inspect at all, only audio, which is where a lot of impersonation fraud now happens.
The same limits apply. A detector can score audio for signs of synthesis and return a probability, useful for triage but not a verdict, and quality drops once a clip is compressed by a call or a messaging app. For a voice or video you actually need to trust, from a remote hire, a claimant or a customer, the stronger move is the same as for images: have it captured live and sealed, so the recording cannot be a clip generated in advance.
When you need to trust a video you did not film
Detection looks backward at a finished file and estimates. The stronger move, when the video is something you can request, is to remove the doubt at the source. If you need a clip from a customer, a claimant or a remote worker, have them capture it live instead of uploading a file. Danaya seals a video the moment it is recorded: it binds a SHA-256 fingerprint of the file to a server timestamp and a GPS position, capture is live only so a pre-made deepfake cannot be imported, and anyone can confirm the result from a public link.
Be honest about the limit. This proves the clip was filmed live at that moment and has not changed since, verifiable by anyone. It does not read the content for truth, so a sealed capture tells you the video is a genuine live recording, not that the scene in it is honest, and the GPS is declared by the device. We make no claim that the result is unforgeable.