Key figure: Deloitte projects US generative-AI-enabled fraud losses to reach $40 billion by 2027, up from $12.3 billion in 2023 (Deloitte Center for Financial Services).
Can you tell a real video from an AI-generated one? Spot the AI Video Duel shows two clips side by side and asks you to pick the fake — a hands-on way to train the deepfake detection skills that AI agents 2026 demand. Under the EU AI Act, AI-generated content must be labeled, and this game shows why that rule exists: synthetic video is now good enough to fool most viewers.
AI agents increasingly act on video content — screening footage, verifying identities, moderating uploads. An agent that cannot distinguish synthetic video from real footage can be fooled by fabricated evidence, fake endorsements, or manipulated instructions embedded in media. Detection skill is a core trust control: before an agent acts on what it sees, something must verify the pixels are real. This game trains exactly that instinct in humans first.
The EU AI Act requires providers of AI systems that generate synthetic content — including deepfake video — to label it as artificially generated in a machine-readable way, so viewers and downstream systems know what they are looking at.
Two clips. One is AI-generated. One is real camera footage. You get about 10 seconds to decide — then instant right-or-wrong, a final verdict, and a shareable score card.
For fun and education — spotting tells like warping backgrounds, too-smooth motion, or physics that feels "off". Not a forensic detection tool.
Every clip's origin, shown on this page after each guess — nothing presented as real that isn't:
None of these are proof — they're tells. Real footage can look strange, and AI keeps getting better. That's the game.
This game reads its rounds from videos/duel/rounds.json. To add your own AI vs real rounds (e.g. clips from Veo 3.1, Kling or Seedance):
ffmpeg -i input.mp4 -vf scale=640:-2 -c:v libx264 -crf 28 -preset fast -an -movflags +faststart -t 10 videos/duel/my-clip.mp4videos/duel/ next to the existing ai-*.mp4 clips.videos/duel/rounds.json — each entry is:{"subject":"…","ai":"videos/duel/ai-x.mp4","real":"videos/duel/real-x.mp4","ai_credit":"…","real_credit":"…","real_page":"https://…","practice":false}ai_credit must say the clip is AI-generated; real_credit must name the real footage source + license. Never mark an AI clip as real."practice": true and note "real_note" — the round is then clearly labeled "practice" and unscored.rounds.json on the next page load.Does AI-generated content have to be labelled under the EU AI Act?
Yes, in most cases. Deployers of AI-generated image, audio or video must disclose it as AI-generated — the Article 50(4) deployer disclosure duty has applied since 2 August 2026 — and providers must ensure AI outputs are marked so they are machine-readable and detectable.
What are the labelling requirements for AI-generated content?
The notice must be clear and distinguishable, and given at first exposure. Providers must mark AI-generated content so it is machine-readable and detectable, with a phase-in deadline of 2 December 2026 for systems already on the market.
How do I label my content as AI-generated?
Put a visible notice on the content itself — not buried in a caption — at first exposure, and add machine-readable provenance such as C2PA Content Credentials or a watermark. Keep a dated record of every labelled asset.
Each round shows two short video clips side by side for about 10 seconds. One clip is AI-generated, one is real camera footage. You tap which one you think is AI, get instant right-or-wrong feedback, and a score at the end with a shareable verdict card.
Yes. The two duel rounds pair an AI-generated clip against real, freely-licensed camera footage, with the source and license credited under each clip after you guess. The two practice rounds use two AI-generated clips and are clearly labeled as practice — they are not scored.
Yes. The page reads its rounds from videos/duel/rounds.json. Drop your own AI and real MP4 clips into videos/duel/, update the JSON fields (ai, real, ai_credit, real_credit), and the game picks them up automatically.
Yes, in most cases. Deployers of AI-generated image, audio or video must disclose it as AI-generated — the Article 50(4) deployer disclosure duty has applied since 2 August 2026 — and providers must ensure AI outputs are marked so they are machine-readable and detectable.
The notice must be clear and distinguishable, and given at first exposure. Providers must mark AI-generated content so it is machine-readable and detectable, with a phase-in deadline of 2 December 2026 for systems already on the market.
Put a visible notice on the content itself — not buried in a caption — at first exposure, and add machine-readable provenance such as C2PA Content Credentials or a watermark. Keep a dated record of every labelled asset.