AI - Anthropic's Claude now embeds an invisible watermark into every piece of text it generates
Anthropic’s Claude now embeds an invisible watermark into every piece of text it generates. This is what is making the rounds today 🤣
Hardly a surprise, given it is not the first in the industry to do so 😊
I do not want every day to be about AI, but I think this is an important development that it will be fun to look at a few years down the line how it all evolved 🤣
AI-generated text is becoming traceable
Anthropic is adding an imperceptible, machine-readable watermark to text from supported Claude models. It is embedded during generation, survives copy-and-paste, and may remain detectable after light editing. The rollout applies worldwide, including Claude accessed through APIs and cloud platforms.
The watermark is not definitive proof of AI authorship. A document edited, translated or proofread by Claude could receive a mark, while substantial rewriting, translation or mixing with other text may weaken or remove it.
This is bigger than spotting AI-written homework. Watermarks and signed provenance could help publishers, social platforms and security teams investigate content origin, but an absent watermark cannot prove that something was written by a human. Everyone is or will be at some point using AI to assist them with whatever it is they do, so I think the sheer existence of said watermark will not be a condemnation in and of itself 😉 I am definitely in favour of adding something that helps us “track” it down though 😊
Google Gemini has already been doing something similar. Google introduced SynthID watermarks for text generated through the Gemini app and web experience in 2024. SynthID subtly adjusts token-selection probabilities without visibly changing the resulting text.
OpenAI has researched text watermarking but has highlighted its weaknesses. It reported that translation, wholesale rewriting and other transformations could defeat its experimental approach, while false positives could disproportionately affect groups such as non-native English writers. Its current published provenance programme focuses on SynthID and C2PA for supported images and audio, not general ChatGPT text.
The real takeaway: AI detection is moving away from unreliable “does this sound like AI?” classifiers and towards model-level provenance signals, but those signals remain evidence rather than proof.
I found out about it on X from here, so giving credit where credit is due 😊