Can we trust what we see?
As synthetic media spreads, telling an authentic recording from an AI-generated one is a serious challenge. Watermarks and provenance data help — but they can be stripped or degraded.
And here's the catch that trips up even careful people: a marker only tells you where content came from. It never tells you whether it's true.
A marker signals origin, not truth. Detection tools tell you something was AI-made — not whether what it shows actually happened.
What you'll be able to do
- 1Compare watermarks and provenance metadata as detection approaches.
- 2Explain their practical limits — stripping and degradation.
- 3State why a marker doesn't prove truth.
- 4Distinguish provider vs deployer duties under the EU AI Act.
- 5Run a verification checklist on a piece of media.
Two ways to tag synthetic media
Invisible watermarks
A signal hidden inside the pixels or audio, designed to survive editing. Travels with the file itself.
Provenance metadata
Content Credentials log a verifiable creation history — signed records of how the media was made.
Why markers fail in the wild
But the deeper limit isn't technical. Even a perfect, intact marker answers only one question — was this AI-made? — and stays silent on another: is it true?
The marker tells you one thing
Click each cell. A marker only reports origin (AI-made or not) — never whether the content is factually true.
Open all four cells.
A video carries a valid AI watermark. What does that establish?
Who has to do what
Providers (Article 50)
Must ensure generative outputs carry machine-readable marks — except for standard assistive editing.
Deployers
Publishing deepfakes or unreviewed public-interest text requires clear public disclosure.
Provider duty, or deployer duty?
The provider builds the system; the deployer uses it to publish. Sort each obligation.
⚙ Provider
📣 Deployer
Drag a chip into a bucket — or tap a chip, then tap a bucket.
Under Article 50, ensuring outputs carry machine-readable marks is primarily whose duty?
Run the checklist
Because no single marker is enough, verification is a process. Run each check on a suspicious clip.
A clip has no metadata at all. The best conclusion is…
Knowledge-check score
Watermark — a signal hidden inside media to mark its origin. Content Credentials / provenance — signed metadata recording how media was made. Provider — builds the AI system. Deployer — uses it to publish content.