Frequently asked questions

It's the detector's score for how much the face looks like the edited faces it learned from. It isn't how much of the video was changed.

Compare it with the line (the threshold). From the line up, the result says likely deepfake. Below it, likely real. A score close to the line means the answer is uncertain.

Each detector has its own line, set when it was tested, so the same score can mean different things. Vidi's line is low, at 23%, so a score of 40% already counts as likely deepfake there. Vici's line is high, at 83%.

No. It means the video looks like the fakes the detector learned from. Heavy compression, strange lighting, a partly hidden face or video filters can make a real video score high.

No. It only means the detector didn't find enough warning signs in the face. It doesn't check the voice, who posted the video, when it was made, or whether it's being shown in a misleading way.

It highlights the parts of the face that mattered most to the result, in the moments that stood out. Brighter means it mattered more. A bright spot alone doesn't prove an edit.

The row of pictures shows the moments the detector looked at most closely. Under each one are the seconds it covers and that moment's own score. A red border and a flag mean the moment rose above the line. Click a picture to see where on the face the detector looked, frame by frame.

The chart below shows the score across the whole video as a curve, with the dashed line for the threshold. The numbered markers are the same moments as the pictures, drawn at their own scores. A single moment can sit above or below the curve, because the curve averages every check around that point in time.

The detector also scores three short clips from the video, the way it was tested in our study. It looks at less of the video than the main result, so the two can disagree. When they do, the main result, which uses the whole video, is the one to go by, and it's worth checking the video with another option too.

One video file (MP4, MOV, WebM or MKV), between 10 seconds and 2 minutes long, up to 150 MB, and between 360p and 1080p. It works best when one person's face is clearly visible for most of the video.

Pick the one that sounds most like your video: clear and good quality, forwarded or re-uploaded, almost too perfect, or dark and shaky. If you're not sure, keep the first option. The Models page explains each one.

Most checks take about a minute. If nobody has used OriginAI for a while, its server needs to wake up first, which can take 5 to 10 minutes. Keep the page open and your result will appear. If you reload the page by accident, it picks the check back up.

Usually because it's too long, too short, too large, below 360p or above 1080p, or because no face could be followed through enough of the video. The page tells you which. Try a clip where one face stays in view, or save it again as an MP4.

Yes, for public videos from the platforms listed on the Scan a video page: right now TikTok, Facebook (including Reels and links copied with Share), X (Twitter) and Reddit. Choose Paste a link, paste the video's link, and OriginAI fetches it for you.

Links from YouTube, Instagram, Vimeo and Dailymotion can't be checked yet. Private, login-only and live videos can't be checked from a link either, and some platforms block it from time to time. In those cases, download the video and upload the file instead. Only check videos you're allowed to use.

It's sent securely to our server, which passes it to a rented computer that runs the detector. The file is deleted as soon as the check finishes. Your result and its map are kept for up to an hour so the page can show them, then deleted.

No. There are no accounts, and we never ask for your name or email to scan a video.

No. Uploaded videos are never used to train or improve the detectors.

Four BS Computer Science students at the University of the East, Manila, as their thesis. Meet the team.

Not on its own. For anything legal, work-related, academic, medical or financial, use OriginAI as one clue among several and ask a qualified person to review the original video.