# Did you know Claude's text will soon carry an invisible watermark?

Hi everyone, it's me! Today I found an announcement that's equal parts fascinating and a little sneaky: Claude's writing is about to get an invisible watermark built right in.

## What was announced?

On Anthropic's News, Anthropic announced that **future Claude models will generate text containing a watermark**. A watermark is a way to determine, after the fact, the likelihood that a given piece of text was written by Claude. Anthropic isn't alone here either — several other major AI model providers are rolling out their own watermarks too.

## Why it matters

Starting August 2, 2026, the EU requires AI providers serving its market to mark AI-generated content. Anthropic, along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content back in July 2026. This watermark is how Anthropic is meeting that requirement. Being able to check after the fact whether AI was involved in producing a piece of text is becoming an industry-wide expectation.

## What changes

The big question is probably "will this change how Claude writes?" — and the answer is no. In Anthropic's internal testing, watermarking showed no impact on the content, creativity, or readability of Claude's output. To a reader, watermarked and unwatermarked text look identical.

Here's what else is good to know:

- No extra tokens are produced, so pricing stays exactly the same
- The impact on generation speed is negligible
- The watermark carries no identifying information — it can't be traced to a specific person, organization, or chat

This applies to Claude models released on or after August 2, 2026, and Anthropic plans to roll it out to older models over the coming months too. Because there isn't yet a reliable way to scope it by region, it's being applied globally at launch. A watermark detection API is also coming soon.

## Dive Deep

The mechanism itself is pretty neat. Large language models like Claude generate text one word at a time, and often there are multiple candidate words that are basically equally good (think "The weather today was cold and overcast" versus "...and grey" — either works fine). Normally, which one gets picked comes down to a random number. With watermarking, the words chosen don't change — only the **source of that randomness** does: instead of an arbitrary random number generator, it's derived from Anthropic's key. Afterward, checking the sequence of word choices against that key lets you estimate the probability that Claude produced the text.

The method is based on Google DeepMind's **SynthID-Text**, published in Nature in 2024, which itself traces back to a 2022 proposal by Scott Aaronson.

That said, watermarking has some real limits.

- **Short passages**: fewer word choices mean less signal, so detection confidence drops
- **Factual content**: sentences with essentially one correct answer (like naming a specific work or fact) leave the watermark little room to act on
- **Light editing or proofreading**: if most of the original words survive, there's little that Claude actually chose, so there's less to detect
- **Code**: places where an exact answer is required (or the code would break) don't get watermarked, though comments with more freedom of phrasing can carry it
- **Heavy rewrites**: replacing essentially every word removes the watermark — though at that point it's arguable whether the text still counts as "AI-generated" anyway

Files like images and SVGs work differently. When Claude creates or edits a supported file type, it attaches a cryptographically signed content credential to the file's metadata, using the open **C2PA** standard — the same one camera makers and photo-editing software use. The file itself isn't altered.

It's also worth noting this is a different approach from AI-detection tools like Pangram, which look for stylistic "tells" in phrasing. Anthropic's watermark instead relies on a cryptographic key, which is a fundamentally different kind of check.

## Wrap-up

- Claude models released on or after August 2, 2026 will produce watermarked text; older models will get it over the coming months
- The goal is compliance with the EU AI Act and its Code of Practice — no impact on quality, speed, or pricing
- Watermarking is weaker on short text, factual content, light editing, and code, and disappears under heavy rewrites
- A watermark detection API is coming soon
- Images and other files get a separate C2PA content credential rather than a watermark

Worth a read if you're setting AI usage policy at your organization, or you're an engineer or content lead who needs to think about how AI-generated content gets handled!