AI Video Detector

Dùng thương mại được. 380+ mẫu Không dấu nước Không cần đăng ký
Mô hình:
+ GPT-5, Claude, Gemini
Upload a clip and we sample a handful of frames, checking each for the tells of AI generation and face-swapping — warped features, melting edges, flickering detail, impossible physics. You get an overall verdict plus a per-frame breakdown. Honest note: modern video models (Sora, Veo, Kling, Runway) are hard to catch and skilled deepfakes harder still. Treat this as one signal, never as proof.

Drop a video to analyze

MP4, WebM, MOV up to 50MB — shorter clips score faster

~3,000 tokens
Sampling frames… 0 / 0

-

-
0%

-

Signals across frames:
    Per-frame scores:

    This is an AI-powered estimate from sampled frames, not a court-admissible forensic test. False positives and false negatives happen — use judgment alongside it.

    When to check a video

    Newsroom verification

    Vet a viral clip before you republish it. Pair frame-sampling with reverse-image-search on keyframes and InVID metadata checks. One signal among several, never proof on its own.

    Deepfake / face-swap

    Check whether a face in a video was swapped — look for boundary flicker around the jaw, mismatched skin tone, teeth that warp between frames. Switch the mode chip to Deepfake for a face-focused pass.

    Catfish / romance scams

    A short selfie video is harder to fake than a photo, but AI avatars exist. Sample it before trusting a profile. A live, unscripted video call is still the gold standard.

    Education / submissions

    Screen submitted video assignments for AI generation before a human review. Tells you a clip is likely synthetic; cannot prove which tool made it. Never the sole basis for an integrity charge.

    Honest about limits

    Top-tier models (Sora 2, Veo 3, Kling 2, Runway Gen-4) produce clips that pass frame-level checks. Re-encoded, cropped, or short clips are harder to score. C2PA content credentials, where present, are more reliable than visual heuristics.

    How it works

    Everything runs in your browser up to the frame check: we grab evenly-spaced frames from your clip, send only those still images to our vision model, and average the results. The video itself never leaves your device as a file — only the sampled frames are analyzed.

    Tùy chọn cao cấp
    Kết quả
    Tiền hết rồi. Lấy thêm biểu tượng
    Muốn kết quả tốt hơn? Các mẫu cao cấp (GPT-5, Claude, Gemini) cung cấp chất lượng cao hơn. Xem kế hoạch

    ❤️ Love this tool? Share it!

    Đăng ký để nhận được một liên kết giới thiệu và kiếm được 25.000 token cho mỗi người bạn.

    Muốn thêm nữa không? Đăng ký miễn phí 30K token/ngày + 10K thưởng
    Đăng ký miễn phí

    Đang xử lý yêu cầu của bạn...

    Upload a clip and find out whether it was AI-generated or deepfaked. We sample frames and check each for diffusion artifacts and face-swap tells. Free, no account.

    Cách sử dụng AI Video Detector

    1
    Nhập thông tin

    Nhập văn bản, tải lên tập tin, hoặc mô tả điều bạn muốn. Không cần tài khoản.

    2
    Nhấn để tạo

    AI của chúng tôi xử lý yêu cầu của bạn trong vài giây sử dụng các mô hình nguồn mở tốt nhất.

    3
    Tải về và chia sẻ

    Tải xuống, sao chép, hoặc chia sẻ kết quả của bạn. Miễn phí cho sử dụng cá nhân và thương mại.

    Dùng công cụ này qua API

    Tự động hóa công cụ này từ mã của bạn. Điểm kết thúc REST tương thích với OpenAI, xác thực bằng token của người mang, không cần SDK bổ sung. Chi phí token tương ứng với giao diện web.

    curl -X POST https://api.free.ai/v1/video/generate/ \
      -H "Authorization: Bearer sk-free-..." \
      -H "Content-Type: application/json" \
      -d '{"prompt": "A cat playing piano", "duration": 4}'

    AI Video Detector — FAQ

    Upload a clip and we sample a handful of evenly-spaced frames, run each through our vision model, and combine the results into one verdict — a probability score, a confidence badge, the signals spotted across frames, and a per-frame score strip. Two modes: AI-generated video detection and deepfake face-swap detection.

    No. Frame sampling happens entirely in your browser: we grab still frames from the clip and send only those images to the detector. The video file itself never leaves your device. That keeps it fast and private — only the sampled frames are analyzed, then discarded.

    Honest answer: modest. It does well on older or lower-effort AI video and obvious face-swaps, but the current top models (Sora 2, Veo 3, Kling 2, Runway Gen-4) produce clips that frequently pass a frame-level check. Treat the score as one signal, never as proof. False positives and false negatives both happen.

    Because it analyzes individual frames as images, not motion over time. A single AI frame can look photoreal even when the temporal behavior (flicker, morphing, unstable detail between frames) gives it away to a human. We surface a per-frame breakdown so you can eyeball consistency yourself, but dedicated temporal forensics goes deeper than what runs in a browser.

    Per frame: warped or melting edges, smeared fine detail, garbled text and logos, impossible reflections and shadows, plastic-looking skin, extra or fused fingers, and diffusion-style texture artifacts. In deepfake mode it weights face-boundary flicker, mismatched skin tone at the jaw/hairline, and teeth or eyes that warp between frames.

    AI-generated video checks whether the whole clip was synthesized by a text-to-video model. Deepfake face swap checks whether a real video had a different face composited onto it. The tells differ — switch to Deepfake mode when you suspect a swapped face on otherwise-real footage.

    No. Courts require provenance-based verification — C2PA content credentials, cryptographic signing, original-source chain-of-custody. Visual heuristic detection has too high a false-positive rate to be admissible. Use this for a casual sanity check and pair it with proper forensics for anything high-stakes.

    Fewer frames means fewer samples to average, so the verdict is shakier — the confidence badge drops to low. Heavy re-encoding (social-media re-compression, screen recording) also smooths over the very artifacts the detector relies on, which can hide a synthetic clip or, less often, make a real one look suspicious. Upload the highest-quality original you have.

    The same vision-language model behind our image detector (Qwen 2.5-VL 7B), prompted for forensic-style artifact analysis. Each sampled frame is scored with a structured JSON schema (probability + verdict + confidence + signals), then aggregated across frames.

    Some generators and cameras embed C2PA signatures that declare how a clip was made — that is the more reliable path than visual heuristics. We do not yet read C2PA from video. Where a clip carries credentials, verify them at verify.contentauthenticity.org alongside our check.

    Hive and Sensity are B2B deepfake-detection APIs priced for enterprise; Deepware offers a free scanner focused on face-swaps. Free.ai gives you a free, no-signup browser check across both AI-generation and deepfake modes with a per-frame breakdown. For legal or brand-safety determinations, the dedicated B2B forensics vendors go deeper — we are a transparent first-pass tool.

    The page samples frames client-side and calls /v1/image/describe/ per frame (mode=ai-detect or deepfake-detect). To automate, extract frames yourself (ffmpeg) and POST each to that endpoint, then average ai_probability across frames. Bearer auth, 10K tokens/month free. /api/ has the curl example.

    Đăng ký miễn phí 30,000 token

    Tạo tài khoản miễn phí

    Không cần thẻ tín dụng

    Anh đánh giá công cụ này thế nào?

    Love this tool? Share it!