Deepfake Detector
A deepfake is not a single technology, and spotting one is not a yes or no question. Here is what the term covers, which inconsistencies are worth checking, and how Drubl can assess suspicious media without overclaiming.
The short version
- Drubl estimates whether media shows signs of synthesis. It does not establish who a person is or what they intended.
- Face, motion and sound each offer clues, and each clue can have an innocent explanation.
- Detection is probabilistic. Context and original sources matter as much as any score.
What a deepfake is
Deepfake is a broad label for media in which a person's face, voice or body has been synthesized or altered by machine learning so that they appear to do or say something they did not. It includes face swaps, where one person's face is placed on another's footage, and fully generated clips that depict a person who never sat before a camera. It also includes cloned or altered voices.
It is worth separating it from other manipulation. A clip that has simply been cut to remove context, slowed down or given a misleading caption is misleading but not synthetic. Ordinary visual effects and beauty filters are also not deepfakes. Drubl's classes of Real, AI, CGI, Filtered and Artwork exist partly to keep these distinctions visible.
Facial inconsistencies worth checking
Small inconsistencies around a face can stand out, but many synthetic faces look convincing. Pause the clip and look closely at these areas without treating a single odd detail as proof.
Face boundary
Where the face meets hair, jaw and neck, look for flicker, blur or a faint seam that does not follow the head's movement.
Eyes and teeth
Blinking that is too rare, too regular or oddly timed, gaze that does not track, and teeth that blur into a single bar can all be signs.
Skin and lighting
A face that is lit differently from the neck or the room, or skin that is much smoother than the hands, deserves a second look.
Extreme angles
Profiles, hands crossing the face and fast turns are difficult for synthesis. Strange behaviour at those moments is informative.
Temporal inconsistencies
Temporal means across time. A single frame might be convincing while the sequence is not. Watch for identity that subtly shifts, with a face seeming to change age or shape across a few seconds, for details like moles and earrings that appear and vanish, and for movement that looks slightly too smooth or oddly jerky. Because these effects occur between frames, they are easy to miss at normal speed. Slow the playback or step through frame by frame where you can. The full video guide walks through this kind of inspection in more depth.
Audio and video together
When a clip includes speech, there are additional questions. Does the voice have natural breaths, room tone and variation, or does it sound flat and evenly paced? Do the mouth movements match the sounds, especially for sounds that need closed lips? Does background noise stay consistent across cuts?
These are manual checks, and they are fallible. Poor recording, translation dubbing and bad connections all produce mismatches in genuine footage. Drubl does process audio signals as one input to its estimate, but it does not verify a speaker's identity and cannot guarantee that lip-sync problems will be identified. Treat any audio finding as supporting context.
Why detection is probabilistic, not absolute
Generators and detectors improve in response to one another. A method that works on last season's fakes may miss new ones, and a clean fake can be recompressed until no useful trace remains. In the other direction, real footage that is low quality can trigger false alarms. So the honest output is a degree of confidence, not a declaration.
This is also why Drubl avoids saying a clip is "proven fake". An estimate that leans toward AI is a reason for caution. It is not evidence that a particular person did something, and it should not be used alone to accuse anyone. The same reasoning applies to stills, which is covered on the AI image detector page.
What to do with suspicious media
Start with the source. Find the earliest upload and see whether reputable outlets or the person involved have addressed it. Run the media through Drubl using a supported link for video or an upload for an image. Compare what you find with what you can see for yourself. If the stakes are high, such as a political claim, a financial request or a threat, do not rely on any single tool: contact the person through a channel you already trust, or ask a professional fact-checking organization. For video-specific checks, the AI video detector page explains how links are handled.
Common questions
- Is this a deepfake?
- Drubl can estimate whether an image or a supported video link shows signs of AI generation or manipulation. It cannot confirm identity or intent, and no tool can promise to catch every deepfake.
- Does Drubl detect fake audio?
- Drubl processes audio signals as one input, but it does not verify speakers and should not be treated as a voice-authentication tool.
- Can a real video be flagged?
- It can. Heavy compression, filters and low quality sometimes resemble synthetic patterns, which is why confidence and context matter.
- What should I do if the result is uncertain?
- Look for the original source, check trusted coverage, and be open about what is unverified when you share.
Something looks off?
Not sure? Let Drubl dig. Check an image upload or a supported video link from the home page.