Glossary
AI watermarking
AI watermarking embeds an invisible marker in AI-generated text, images, audio or video so machines can later identify the content as artificial. Google's SynthID and the C2PA content-credentials standard are the best-known approaches. Useful, but not foolproof.
As AI-made images, voices and videos became indistinguishable from real ones, the obvious question followed: how do we tell? Watermarking is the main technical answer. The generator hides a signal in the content itself (patterns in pixels or audio a human can’t perceive) or attaches signed metadata recording how the file was made. Google’s SynthID does the former for images, audio and text from its models; the C2PA “Content Credentials” standard, backed by Adobe, Microsoft, OpenAI and others, does the latter. The EU AI Act pushes in the same direction, requiring AI-generated content to be machine-detectable.
Here’s the honest part, because watermarking is often oversold: it works best when content goes straight from generator to viewer. Cropping, re-encoding, screenshots or running content through another tool can weaken or strip the marks, and detection tools for unwatermarked content are unreliable. So treat a watermark as positive evidence (“this is AI-made”), but never treat its absence as proof of authenticity. For deepfakes aimed at you, verification habits still beat detection tech.
Where you’ll meet this
- “AI info” or “Content Credentials” labels on images on Meta, YouTube and LinkedIn
- Google’s SynthID notices in Gemini-generated images and audio
- Cameras and editing tools (Adobe, Leica) that attach provenance metadata