Anthropic launches new model that imperceptibly watermarks texts made with their Language Models: ‘The load-bearing question is, how would anyone know if it was watermarked?’

Author image for Olivia Arocena
Olivia Arocena • Writer and Editor
Aug 18, 2026 8:00 AM EDT

The trial is officially over

It was about time. You know what they say: when a product is free, it's because you are the product, or something like that. We've enjoyed LLMs very much up to this point, and while we were enjoying them, we also helped them learn how to fix the logical and ethical problems that arose in the process, and this post is about that.

Companies like Anthropic introduced the use of AI so elegantly that it became necessary; so now, we're kinda dependent on them, for better and for worse. This update doesn't change whether you have the paid or the free version; everything produced by it will still be watermarked.

This is a way for the company to comply with transparency regulations that are very relevant nowadays. Of course, most of us can tell when something is produced by an AI, but sometimes we can't, and if its use remained unregulated, it could cause a massive crisis in fields like education, investigation, marketing, and so on.

We haven't seen the end [luckily], in my opinion, of the regulations that will be imposed on Language Learning Models, and we, as users, have a long road ahead. In this post, you will see an overview of everything that's about to change and how it works.

Some trend forecasters, after the AI slop invasion in the media, predict that the next big trend will be revaluing human talent for writing and creative ideas, and that AI use will be limited to administrative or repetitive tasks. I still think we have a few years left of AI-generated music, film, and marketing campaigns to get there, if we ever do.

Advertisement

  • New Watermark Rules

    *Claude
    via Brecht Corbeel

    Claude's new update

    Still frame for contextual reference only

  • How would an “invisible watermark” in AI-generated text actually work?

    I saw that Anthropic is apparently planning to embed invisible watermarks into Claude- generated text, and this part caught my attention:
    via Several_Fly694
  • Does anyone here know how something like this actually works technically? If it's genuinely part of the text, I'm guessing it can't just be hidden Unicode characters, because those would be pretty easy to
    via Several_Fly694
  • detect and strip. So is it more like: choosing certain words/synonyms according to a statistical pattern?
    via Several_Fly694
  • slightly biasing token selection during generation? encoding a signal into the distribution of words or sentence structures?
    via Several_Fly694
  • something else entirely? And how could it survive editing? For example, if someone takes Claude's output, rewrites 20-30% of it,
    via Several_Fly694
  • changes sentence order, or runs it through another LLM, would the watermark still be detectable? I'm not asking about Al detectors in general. I'm specifically curious about the technical mechanism
    via Several_Fly694

  • All the AIs we use

    ΑΙ ChatGPT DeepSeek Claude Mistral Al Gemini Copilot 00 Poe
    via Solen Feyissa

    Can we really predict how AIs will evolve?

    Image depicted solely for explanatory purposes

  • behind a watermark that's supposedly embedded in the text itself and survives copy/paste. Would love to hear from anyone who understands text watermarking or has read the relevant research.
    via Several_Fly694
  • FreakDeckard Here's how it works in simple steps: 1. The Al uses a secret rule when choosing words. Normally, an Al picks the next word based on probabilities. With a watermark, it slightly favors certain words that match a hidden pattern.
    via FreakDeckard
  • 2. This pattern is created using a secret key and the words that came before. The Al gives a small boost to words on a "good" list or with higher scores. 3. The change is so tiny that the text still sounds natural and makes sense. You can't tell by reading it.
    via FreakDeckard
  • 4. To check if text has a watermark, a detector uses the same secret rule to see if the text has more of the favored words than expected. It gives a score (like a z-score) to decide. 5. The watermark stays if you copy-paste the text or make small changes. But if you rewrite the text a lot (heavy paraphrasing), the watermark disappears.
    via FreakDeckard
  • 6. Different Al companies do it slightly differently: Google's Gemini uses a method called "tournament sampling." Claude (by Anthropic) also adds watermarks at the model level, according to their documents.
    via FreakDeckard

  • Human meets Robot

    Cheezburger Image 10656844032
    via Igor Omilaev

    How will our relationship with AI evolve from now on?

    Scene is illustrative and does not depict actual individuals or events.



     

Want More? Follow Us and Add Us as a Preferred Source on Google.

Tags

claude internet work technology ai tech news workplace Tech artificial intelligence
Scroll Down For The Next Article