Jargon:Watermarking
1 min read
Say this at the party
“AI watermarking hides an invisible mark in AI-made pictures, audio or text, so a computer can later tell it came from an AI.”
In plain English
A watermark on a banknote shows it's genuine. An AI watermark works the other way round: it quietly labels something as machine-made.
There are two main approaches. One hides a signal in the content itself, like Google's SynthID. The other attaches a signed record of where a file came from and how it was edited, called Content Credentials (built on a standard called C2PA).
Neither is perfect. Marks can be weakened by heavy editing, and only some companies use them, so no mark doesn't prove a human made it.
Where it came from
Hiding marks in images and audio is decades old. For AI, researchers published a well-known method for watermarking chatbot text in 2023, and Google DeepMind described the SynthID text watermark in Nature in 2024 after testing it on nearly 20 million Gemini replies.
See it in action
Sam sees a striking photo online and drops it into a Content Credentials checker. It shows the image was made with an AI tool and then cropped, so he doesn't share it as real news.
What people get wrong
- A watermark doesn't say whether something is true. It only says where it came from. A real photo can still be misleading.
- No watermark doesn't mean human-made. Many AI tools don't add one, and marks can sometimes be stripped out.
Go deeper
- PaperDathathri et al., Google DeepMind, Nature, 2024Scalable watermarking for identifying large language model outputsHow SynthID marks AI text without hurting quality, tested on live Gemini replies.
- GuideGoogle AI for DevelopersSynthID: Tools for watermarking and detecting LLM-generated TextGoogle's developer guide to adding and detecting SynthID text watermarks.
- GuideC2PAC2PA | Verifying Media Content SourcesThe industry standard behind Content Credentials, explaining how provenance records work.
- PaperKirchenbauer et al., 2023A Watermark for Large Language ModelsA well-known early method for hiding a detectable signal in chatbot text.
- ArticleWikipediaContent CredentialsWhat Content Credentials are, and a fair note on what they can't prove.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.