How watermarking works and why you should care

Starting from August 2026, Claude models generate a text that contains a watermark.

“But what does it even mean?” One of my friends asked me over a coffee.

“Where can I see that in the text? Are there any hidden signatures? And how can I get rid of it?”

My friend, let’s call him John uses Claude on his workplace, and he is a heavy user. So this was a major reason for him to buy me the coffee. Since I was unsuccessful to explain it to him on the napkin over a 30 minute coffee break, this issue is my gift to him (and to you as well).

Ok, so here how where the magic, I mean math happens:

Large language models like Claude work by generating one word at a time. Each time the model decides on the next word, it chooses among a list of potential candidates, ultimately selecting the most sensible or likely based on the preceding text.

Quote from the official Anthropic documentation page.

It splits sentences into tokens (tiny chunks- words, parts of words, or even punctuation).

Then it assigns a number to every word. For Claude a sentence is a sequence of numbers.

So Claude generates one token at the time.

Now comes the fun part: when you give it a sentence, Claude uses conditional probability to figure out what comes next.

You can find out more how LLMs work in this issue: 𝐇𝐨𝐰 𝐝𝐨𝐞𝐬 𝐚 𝐋𝐋𝐌 𝐰𝐨𝐫𝐤?

Let’s understand how it works on a simple example.

What word comes next? The model gives every possible next token a score.

For example, delicious might get a highest score, fantastic might get a lower score, incredible might get even lower score and the word that makes very little sense like funny will get extremely low score. That raw, unnormalised scores are called logits.