Understanding AI Text Watermarking

What you’ll learn and how you can apply it

• Explain how Claude uses the Google SynthID-style text watermarks to hide a statistical signal in word choice, and why detection is impossible without the secret key
• Generate watermarked text with an open weight model, then detect it on an ordinary laptop using mean scores, z-scores, and p-values

Course description

Every major AI provider now watermarks the content its models produce, and the EU AI Act has turned what was once a research curiosity into a legal requirement. But most people, including many engineers who work with these models daily, have never seen a watermark, couldn’t point to one in a sentence, and don’t know what it takes to detect one. That’s the gap this course closes.

In one hour, you’ll explore invisible image and text watermarking. First, compare identical-looking photos, detect a hidden watermark, and see it survive JPEG compression, resizing, and screenshots. Next, learn how text watermarking embeds signals directly within words. Examine Google’s SynthID-Text using secret keys, g-values, and tournament sampling, and observe a detector scoring watermarked versus unwatermarked outputs. Finally, discover why human writing and Claude output return “not detected,” and analyze failure cases like short passages, factual answers, code, and zero-temperature generation. You’ll be able to explain to a colleague, a compliance team, or a skeptical stakeholder what a text watermark is, what it reliably tells you, and what it can’t tell you.