3 AI writing tools, only 2 watermark your text
Three AI writing tools, and only two of them mark the text they hand you. Here is where each one stands as of August 2026.
Gemini. Google has been watermarking Gemini text for a while using SynthID, deployed inside Gemini at production scale. It was the first text watermark of its kind running live at that size. Google has also signed the EU code of practice on transparency and is pushing SynthID into partnerships with other labs.
Claude. Anthropic began embedding a machine readable watermark in text from new Claude models on 2 August 2026. It applies worldwide rather than only in the EU, and covers output from the API, coding tools and cloud partners.
ChatGPT. No text watermark. OpenAI developed the technology and deliberately did not ship it. Two reasons have been reported: surveys suggested people would use the product less if the output could be traced, and there was concern about unfairly stigmatising writers who use it as a legitimate aid, particularly people whose first language is not English. OpenAI does watermark images, just not text.
How these marks work. When the model generates, it picks each word from a range of plausible options. The watermark biases those choices according to a hidden pattern. The text reads completely normally. The signal is spread across the passage and lives in the word selection itself.
This is why it is not hidden characters. There are no zero width spaces to strip, no invisible Unicode to clean. Anything sold as a watermark remover that works by deleting invisible characters is addressing a different problem entirely.
Where the marks are weak. Every provider says the same thing, and the academic work agrees: detection degrades under paraphrasing, translation, heavy editing, and mixing generated text with your own writing. Short passages may not carry enough signal to detect at all.
Worth knowing that no provider has published a public text detector yet, so none of this can be verified from the outside right now.
In my experience using AI writing tools, the presence or absence of text watermarks can have significant implications for content creators and users alike. Google's Gemini, equipped with SynthID watermarking technology, sets a strong industry example by embedding traceable signals into the choice of words itself rather than relying on invisible characters. This approach ensures that while the text remains completely natural and readable, it still carries metadata useful for identifying AI-generated content. I find this particularly helpful when verifying authorship or ensuring transparency in collaborative projects. Anthropic's Claude has also joined the movement by introducing machine-readable watermarks globally starting August 2026. This universal approach across APIs and coding tools shows a commitment to accountability beyond any single region, which could shape wider adoption and standardization in the AI writing space. On the other hand, OpenAI’s ChatGPT does not use text watermarking, despite originally developing the technology. This decision, motivated by concerns over user adoption and protecting legitimate writers, reflects the nuanced balance between technology and user sensitivity, especially for non-native English speakers who may rely on AI for assistance. From my own use, while no watermarking means less traceability, the technology still offers unparalleled aid in drafting and ideation. One interesting technical insight is that these watermarks do not use zero-width spaces or invisible Unicode characters, so commonly alleged "watermark removers" that delete such elements do not affect these watermarks. Instead, the watermark is encoded in subtle biases in word selection probabilities, making it resilient to simple text cleaning but vulnerable to heavy paraphrasing or translation. While detection tools remain non-public and experimental at this stage, acknowledging that short passages or heavily edited AI text can elude detection is crucial. This realism helps manage expectations in professional or academic settings where the provenance of AI-assisted writing might matter. Having explored tools like humantone.io, which aims to rewrite text and remove watermarks by altering phrasing, I recognize both the innovation and challenges in maintaining watermark integrity and authorship verification. As AI writing becomes more pervasive, understanding these watermarking technologies and their limits is invaluable for anyone who relies on or interacts with AI-generated content.