AI watermarking means hiding a signal inside an AI-made picture, video, sound clip or piece of text, so that software can later tell it came from an AI tool. Most of these marks are invisible to people; only a special checker can find them.
Hold a banknote up to the light and a hidden picture appears. You never notice it in everyday use, but a bank can check it in seconds. An AI watermark works the same way: it is built into the content itself and waits for someone with the right tool to look.
In a sentence: “The newsroom checked the photo for an AI watermark, but finding none did not prove a person had taken it.”
What it is
The mark lives inside the content, not in a label beside it. The US standards body NIST says a watermark is encoded “into the content itself”, not carried in a separate place such as metadata.1
There are two kinds. A visible one is like a logo in the corner. An invisible one changes the picture, sound or words so slightly that you cannot tell; NIST calls these “covert”.2
A simple example
Google’s SynthID is one example. Google says it covers four kinds of content.3 Anthropic adds a watermark to text from its newest Claude models.4 For text, the mark is a pattern in which words get picked, not a hidden character; Claude’s text watermark explains how.
Why it matters to you
Some laws now require it. In the European Union, the AI Act says AI providers must mark what their tools generate.5
But a watermark is a clue, not proof.
- Marks can be removed. NIST warns that most current schemes can be stripped out.2
- Rewriting weakens text marks. Google says its own detector’s confidence drops sharply when AI-written text is thoroughly rewritten or translated.3
- No mark proves nothing. Not every tool adds one, and each company’s mark needs its own checker, as Anthropic explains.6 Never treat “no watermark found” as “made by a human”.
Often confused with
| Term | How it differs |
|---|---|
| Content Credentials | Differs: a signed record attached to the file, like a label on a parcel. A watermark is woven into the content itself. See Content Credentials meaning. |
| AI detectors | Differs: a detector guesses from style whether something looks AI-made. A watermark checker looks for a mark that was deliberately put there. |
Questions people ask
Can AI watermarks be removed?
Often, yes. NIST, the US standards body, says “most current watermarking schemes have been consistently found vulnerable to removal”.2
What is SynthID?
It is Google’s AI watermark. Google says it embeds marks “directly into AI-generated images, audio, text or video” and can later identify them.3
Is a watermark the same as Content Credentials?
No. A watermark is hidden in the content itself; Content Credentials are a signed record attached to the file. NIST describes watermark information as encoded into the content “without using a separate channel (e.g., metadata fields)”.1
Learn more on this site
- How AI content gets labelled — watermarks and Content Credentials side by side, and what each misses.
- Claude’s text watermark — what it is, what it can prove and why the EU requires it.
- Deepfake meaning — the fakes that labelling is meant to help catch.
- The AI glossary and AI for beginners — the basics, in plain words.
Related on this site: Do AI Detectors Work?
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is an AI watermark?
- A label printed next to the file name
- A guess about style made by an AI detector
- A hidden signal inside AI-made content
Show the answer
Answer: A hidden signal inside AI-made content. The page says the mark lives inside the content itself, and most are invisible until a special checker looks.
2. A checker finds no watermark. What does that prove?
- Nothing: the mark may be gone or never added
- That a person definitely made it
- That the content is safe to share
Show the answer
Answer: Nothing: the mark may be gone or never added. The page says marks can be removed, not every tool adds one, and each company’s mark needs its own checker, so no mark proves nothing.
3. How are Content Credentials different from a watermark?
- They are woven invisibly into the content
- They are a signed record attached to the file
- They are a guess based on writing style
Show the answer
Answer: They are a signed record attached to the file. The page says a watermark is woven into the content, while Content Credentials are a signed record attached to the file, like a label on a parcel.
Related terms: AI ethics · AI governance · Deepfake · Shadow AI · Slop · Voice cloning
- 1NIST, NIST AI 100-4: Reducing Risks Posed by Synthetic Content, November 2024, read at source 1 Oct 2026“Digital watermarking involves embedding information into content (image, text, audio, or video), typically while making it difficult to remove. The distinguishing characteristic of watermarks is that the information is encoded into the content itself via the pixels, words, etc., without using a separate channel (e.g., metadata fields) to convey the information.” ↑a“Digital watermarking involves embedding information into content (image, text, audio, or video), typically while making it difficult to remove. The distinguishing characteristic of watermarks is that the information is encoded into the content itself via the pixels, words, etc., without using a separate channel (e.g., metadata fields) to convey the information.” ↑b
- 2NIST, NIST AI 100-4, November 2024, read at source 1 Oct 2026“Covert watermarks are machine-readable watermarks involving subtle perturbations of the content that are hard for humans to detect.” ↑a“most current watermarking schemes have been consistently found vulnerable to removal”. ↑b“most current watermarking schemes have been consistently found vulnerable to removal”. ↑c
- 3Google AI for Developers, SynthID, last updated 9 Apr 2025, read at source 1 Oct 2026SynthID “watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text or video.” ↑a“Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language.” ↑bSynthID “watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated images, audio, text or video.” ↑c
- 4Claude Help Center, How Claude marks AI-generated content, read at source 13 Oct 2026“Generated text will carry embedded watermarks”. ↑
- 5EU AI Act, Article 50(2), via artificialintelligenceact.eu, Article 50, read at source 1 Oct 2026outputs must be “marked in a machine-readable format and detectable as artificially generated or manipulated”. ↑
- 6Anthropic, How Claude’s text watermark works, 14 Aug 2026, read at source 22 Sep 2026 (as cited on this site’s provenance page), on another company’s watermark: “it would have a different key; it might also use a different watermarking method altogether”.↑
EU AI Act (Regulation (EU) 2024/1689), Article 50, via artificialintelligenceact.eu, read at source 1 Oct 2026.
The comparisons on this page (the banknote, the parcel label) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK