Hello Reader,

Anthropic announced that future Claude AI models will embed machine-readable invisible watermarks directly into generated text. We all know about watermark on the images - it's some faint logo, text, signature on the image background or an attribute in image metadata. This prevents someone copying and using the an original image as their own.

BUT, how can a text have watermark? Text is just text, and moreover if you copy text from model output to notebook to LinkedIn, how can the watermark persist?

In this edition, we answer that question by going through how text can be watermarked.

The core idea

Secretly split the vocabulary into two random groups - a "green list" (~half the possible words) and a "red list" (the other half). Think of this as synonyms of same word. Example: the word happy can be expressed as joyful, glad, content, cheerful. Secretly, Anthropic will put happy, joyful, and glad in green list, and content, cheerful in red. And when generating text for happy feeling, it will use more green words.

Every word and it's synonym will have this green and red list. Then Anthropic can nudge the model to generate text with more green words. As an user, you have no idea which words are green. But, the detection algo will know, and if the ratio of green vs red goes over a certain number, Anthropic will mark the text as Claude generated.

This is why, even when you copy paste from Claude, to notepad, to LinkedIn, the watermark is still there.

The advanced version

There is a published algorithm named SynthID, where each word and synonyms are assigned a g-value, which can be between 0 and 1. So, for the above example, the word and the g-values are: happy 0.71, joyful 0.65, glad 0.63, content 0.25, cheerful 0.32. Basically, the green words have g-value more than 0.5, red words less than 0.5

Now imagine, each word and it's synonyms with g-values.

Then, the LLM can generate the whole paragraph with the target of average g-value of more than 0.5. A detection algorithm can simply derive the average g-value of the whole text, and if it's above 0.5, it'd be deemed AI generated.

Is there a way to avoid it?

The obvious question you are thinking is, how to use AI, but not get detected. Below are the considerations:

• Obviously, if you know which words are green, you can just swap those words with red. But, assuming Anthropic will just expose the Algorithm via an API check instead of publishing the list, so popular platforms (e.g. LinkedIn) can check if your post is Claude generated, and mark it as such.
• Anthropic has committed to supporting detection but hasn't published the actual tooling. So right now, nobody outside Anthropic can actually check whether a given piece of text carries the mark.
• The API will either be gated with some cost, or have some detection if someone is trying to extract the list of green words.
• You can change enough words, so it may go below the average g-value of 0.5. However, it may be a wild goose chase, or a lot of trial and error.

My personal input

I actually like this announcement. I hate seeing AI-slop everywhere. Students tell me that they like my newsletters, especially the interview ones, and explaining the technical concept ones. It's funny because, all of those I type myself. Hence there are always typos and grammar error, but you get the true teaching from me.

With this AI watermarking, I hope the number of AI slops reduce, and it also reduces the pressure from me to generate insane amount of social medial content using AI, which I hate doing.

However, I do believe, this will reduce LLM usage, and I am certain frontier labs don't want that. It's yet to be seen what final solution they come up with. But pretty sure, they'd have a way for you to pay more and get some bypass. I am curious to see what Anthropic announces. But that's a topic for another day in the future.

Now you understand how text watermarking works. I don't believe this will be asked in an interview, however you can impress your friends and co-workers during the next AI debate over a drink!

Keep learning and keep rocking 🚀,

Raj

P.S. If you have found this newsletter helpful, and want to dive deeper...

Everything I teach on AWS, System Design, Kubernetes, and DevOps - all in one Skool classroom. Build the technical depth at your own pace, and get every new course I release: Join Here​

🎥 Free tutorials on Cloud Gen AI + interview prep on my YouTube channel: Here​

🚀 The SA Bootcamp is the full transformation - hands-on practice through Cloud Colosseum, interview prep, mock interviews, and resume reviews, and so much more...
Everything you need to land the $200k+ Solutions Architect role.

​Join the waitlist to be first in when the next cohort opens.