Synthetic Media Provenance Kit
Donnez à la vidéo et à l'audio générés par IA une origine vérifiable.
Un kit pratique pour marquer la vidéo et l'audio générés par IA avec filigranes, métadonnées et mentions de divulgation — ancré sur SynthID, le système de filigrane que Google déclare embarqué dans toutes les sorties de Genie 4.
Installez-la en 3 étapes
Sans inscription. Sans outils en ligne de commande. Choisissez votre IA, copiez une ligne, collez-la.
Vous préférez le faire vous-même ? Achetez la compétence ($4.99), téléchargez le ZIP depuis votre confirmation d'achat, décompressez-le dans le répertoire indiqué ci-dessus, redémarrez votre session IA — même résultat. Vous l'avez déjà achetée ? Votre assistant IA peut la télécharger pour vous.
Contenu du pack
Un fichier SKILL.md (le guide pratique) + un modèle de départ. Aperçu :
Synthetic Media Provenance Kit
The idea in one paragraph
When a world model can generate two minutes of photorealistic video with native audio in real time, the question stops being "can it fool someone" and starts being "can we prove where it came from." This skill is a provenance kit: a repeatable checklist for attaching watermarks, metadata, and plain-language disclosures to synthetic video and audio, anchored on SynthID — the watermark Google states all Genie 4 generations carry.
What it is built on
- Per DeepMind's October 27, 2026 announcement, all generated video and audio from Genie 4 carry SynthID watermarks.
- Genie 4 is described as the first world model with natively generated real-time audio — ambience, music, and interactive dialogue — per Google's announcement; reporting notes (datanorth.ai) that Runway's earlier GWM Worlds 2 also ships playable worlds with audio, so the "first" claim is Google's own framing, not settled fact.
- Genie 4 is a research preview for trusted testers with a waitlist; access is planned for Google AI Ultra subscribers (inside Project Genie) and early Gemini API / AI Studio partners. No dates, prices, or published benchmarks were announced.
Workflow
- Inventory the asset (input: media file -> output: asset record). Note the generator, model name and date, prompt or seed, and who produced it. Provenance starts with knowing what you have.
- Verify the built-in watermark (input: asset record -> output: watermark status). Check whether the platform's native watermark (e.g., SynthID for Genie 4 outputs) is present and intact. Record the check; do not assume it survived editing.
- Attach portable metadata (input: asset -> output: tagged file). Add creator, generation date, model, and prompt summary into the file's metadata or an accompanying manifest. Metadata travels where watermarks sometimes do not.
- Write the disclosure line (input: asset record -> output: disclosure text). Draft one plain sentence: what is AI-generated and what is not. Put it where the audience will actually see it — caption, description, or first comment.
- Lock the chain (input: tagged file + disclosure -> output: publish package). Keep an unedited original alongside the published version, and record both hashes or IDs. If someone questions the asset, you can show the chain.
- Recheck after edits (input: edited version -> output: updated package). Every cut, re-encode, or re-upload can strip watermarks and metadata. Re-verify after any edit and update the record.
What good looks like
Aperçu du modèle
Synthetic Media Provenance Checklist
Disclosure line (one sentence, plain language):
Asset record:
- Generator:
- Model and date:
- Prompt / seed summary:
- Produced by:
Le pack de compétence lui-même (SKILL.md + modèle) est rédigé en anglais.
