the short version
- →most AI detectors handle text only. amige. is one of the few consumer tools that checks video at all.
- →two cases read differently to a detector: fully AI-generated video (Sora, Veo, Runway) and deepfakes, where a real face is swapped onto a real clip.
- →deepfake detection runs harder in the wild than in benchmarks, and social-platform compression adds to that, so amige. shows a range.
- →free to start, and it reads any C2PA / SynthID provenance the file still carries.
how to spot an AI or deepfake video
useful tells, with the same caveat as ever, since the next model release trains them away:
- →face-swap seams. shimmer or a soft boundary at the hairline, jaw and ears; skin tone that shifts where the swapped face meets the real head.
- →eyes and mouth. irregular or too-regular blinking, teeth that smear in motion, and lip movements that drift out of sync with the audio.
- →flicker between frames. details that pop or wobble frame to frame (earrings, background text, fingers), because the model builds frames more independently than a camera captures them.
- →physics over time. hair, cloth and water that move oddly; lighting and shadows that shift as the subject turns.
- →audio. a too-clean or faintly robotic voice, or ambient sound that doesn’t match the scene.
on a short, compressed social clip these get hard to see, which is why a detector that reads frame-level signals beats the naked eye, and why amige. reads several at once.
what amige. checks
amige. samples frames from the clip and runs them through a panel of independent video detectors, including Hive, AI or Not, Sightengine, BitMind and Reality Defender, which together cover both AI-generation and deepfake face-swap signals, then shows you where they agree and where they split. you also get a best guess at the most likely model and a read of any C2PA / SynthID provenance in the file.
the free daily tier works on video. the Pass widens the panel and the limits. the full method sits on how it works, and what a deepfake is sits in the glossary.
where video detection falls short
- →the wild runs harder than the lab. independent evaluations put deepfake detectors below their benchmark numbers on real-world video.
- →compression erodes the signal. social platforms re-encode and compress video, which strips metadata and weakens frame-level signals. the original file gives the better scan.
- →the newest models are hardest. the latest generative-video releases are tuned for photorealism. a confident call on current-gen video carries weight without closing the case.
- →false positives exist. heavy filtering, beauty effects and aggressive compression can make real video read as synthetic. one number won’t carry a serious decision.
questions
how do I check if a video is AI-generated?
upload the clip to an AI video detector. amige. samples frames and runs them through several independent detectors built for AI-generated and face-swapped (deepfake) video, then shows you where they agree and where they split, and reads any embedded provenance. read the spread of detectors over a single number.
can a deepfake be detected?
often, though not always, and a trustworthy tool says so. detectors look for face-swap seams, irregular blinking and lip-sync drift, flicker between frames, and lighting that doesn’t hold up. independent evaluations put real-world deepfake detection below the lab benchmarks, and it weakens further on compressed, re-encoded social-platform video. a panel of detectors plus a confidence range carries more weight than one confident verdict.
is there a free AI video detector?
amige. has a free daily scan tier that works on video, not only images or text. it suits a gut-check on a clip. for anything that carries weight, read the per-detector breakdown over a single free score.
what’s the most accurate deepfake detection tool?
no single tool wins everywhere. accuracy depends on the model that made the video, the compression, and whether it’s a full AI generation or a face-swap. amige. runs several independent detectors at once and shows the disagreement, rather than claiming one ‘most accurate’ number. read a single advertised accuracy figure with caution.
is making a deepfake illegal in the US?
it depends on the use and the state, and this isn’t legal advice. several US states restrict non-consensual intimate deepfakes and election-related deepfakes, and federal rules have moved the same way, while a clear parody may fall outside them. detection tells you whether a video is likely synthetic; whether a specific use breaks the law is a separate, jurisdiction-specific question.
can amige. detect Sora, Veo or Runway videos?
the video panel aims squarely at those generative-video models, and amige. keeps a per-model guide for each. as with images, the newest releases are the hardest to flag and accuracy drops on them, so amige. shows a confidence range on borderline clips.
sources.
- 01C2PA Content Credentials — technical white paper (2025)the provenance standard amige. reads on video files; also covers why metadata gets stripped on re-upload.
- 02Google — SynthID detection built into Chrome and SearchSynthID now spans image, audio and video for participating models.