On August 11, 2026, I discovered that two of the three AI models I use daily leave an invisible signature in the text they generate. Claude, according to its published support documentation, inserts an imperceptible watermark in text. Gemini uses a system called SynthID that biases token selection at the probabilistic level. GPT does not mark the text it generates.
I have been building a complete trilingual blog with these tools for 6 months. Every post was generated in a conversation with an AI and then extensively edited by me. I was never asked whether I wanted the text watermarked. I was never informed it was happening. I found out by reading technical documentation. What follows was verified on August 11, 2026. Watermarking systems evolve quickly and some details may have changed after that date.
How the watermark works
It is not hidden characters. It is not metadata inserted in the text. It is not something you can see, copy, or delete by editing the document.
The system operates at the word selection level. When a model generates text, it chooses the next word based on probabilities. "My favourite tropical fruits are..." and the model calculates: "mango" has 50% probability, "lychee" 30%, "papaya" 15%, "durian" 5%. Without a watermark, it picks "mango." With a watermark, the system biases the probabilities using a secret key: "mango" drops to 45%, "durian" rises to 20%. The chosen word can change. The final text can be different from what the model would have generated without the mark.
This is not an analogy. According to SynthID's technical documentation and the description Gemini provided of its own mechanism, this is how the system operates. The mark exists in the statistical pattern of word selection, not in the visible characters.
Who marks and who does not
I asked each model directly whether it marks the text it generates. I used specific prompts asking what exists in production today, not what has been researched or planned.
Claude marks text with an imperceptible watermark and generated files (SVG, PNG, JPG) with signed C2PA metadata. The text mark persists through copy and paste and may degrade with editing.
Gemini marks text with SynthID, a system that biases token selection using a private mathematical key. It is active in the Gemini app, Gemini Advanced, and API pipelines. It survives copy-paste and minor edits. It degrades significantly with heavy paraphrasing or translation.
GPT, according to available documentation and the model's direct response when I asked, has no documented watermarking mechanism for plain text in production. OpenAI deployed SynthID and C2PA for images, and SynthID for audio.
The auto-attribution toll
If I have difficulty writing and dictate a text to an AI to help me structure it, the watermark does not distinguish between my case and someone who asked the AI to invent everything from scratch. The mark says "processed by AI" for both. It does not say "dictated by a human with writing difficulty and transcribed by AI." It does not say "draft generated by AI and extensively edited for an hour by a human."
It is as if Word claimed authorship for correcting your spelling. Or if the printing press claimed ownership for reproducing your text. The tool left a mark on the output. But the tool's mark is not the authorship of the work.
And there is an accessibility dimension nobody is discussing. People who use AI as an accessibility tool for writing (people with dyslexia, people writing in a second language, people with motor difficulties) receive the same mark as someone who generated mass spam content. The mark does not distinguish. To the detector, both are "AI-generated."
The Ship of Theseus
If you edit every sentence of an AI-generated text, is it still AI text? The watermark says yes (if it survives the editing). The editorial process says no (if every decision about the final text was the human's).
My process is: Claude generates a draft. I read it completely. I question the decisions. I verify the data. I adjust the tone. I rewrite sections. I remove what is unnecessary. I add what only I can add. The text I publish is not the text Claude generated. It is the text I decided deserved to be published after an hour of editorial work.
But the mark does not know that. The mark only knows that at some point token selection probabilities were biased. It does not know whether the human then changed everything, something, or nothing.
The marks are invisible to each other
When I asked Gemini whether it can see Claude's watermark in a text, the answer was clear: no. Each company uses its own secret key. To Gemini, Claude's text is plain text. To Claude, Gemini's text is plain text. The marks are invisible to everyone except whoever placed them.
And when you return edited text to the same AI, the previous mark breaks. The AI generates a completely new mark on the new text. There is no accumulation of marks. No layers. Each generation creates a clean signature from scratch.
Public detectors (GPTZero, CopyLeaks) do not read SynthID or Claude's watermark. They use their own predictive models with significant false positive rates. There is no universal text AI verifier. There is no global registry. Verification of these marks depends on detection mechanisms specific to each system and there is no universal verifier that reads them all.
Three concepts the industry conflates
Authorship is who contributes the ideas, experience, and intention. Assistance is who helps produce the text. Provenance is how we can demonstrate the process through which it was produced.
The watermark conflates all three as if they were one. It says "AI was involved" without specifying whether the AI was the author, the assistant, or just the transcriber. For a professional blog where every post goes through human conception, mixed research (human + AI), AI editorial assistance, and human final editing, the label "AI-generated" does not describe the process. It simplifies it to the point of distortion.
What I am not going to do
I am not going to use tools to remove watermarks. They already exist. A tool for Claude appeared on GitHub days after the announcement. Passing text through another model breaks the SynthID mark almost completely. But removing the mark does not solve the problem. It circumvents it.
I am also not going to pretend I do not use AI. Every post on this blog was generated with Claude's assistance. That is not a secret. It is the subject of the blog.
What I am going to do
Build my own provenance chain. Not an invisible watermark that says "AI was here" without context. An explicit editorial declaration that says exactly how each piece of content was produced.
Author: Diego Sanchez. Conception: human. Research: human + AI. Editorial assistance: Claude. Final editing: human. Sources: manually verified.
That is more honest than a watermark. More useful for a reader who wants to know how the content was made. And more compatible with the GEO infrastructure I already have: JSON-LD, llms.txt, semantic entities, author metadata.
The watermark says "AI was here" and nothing more. My provenance chain says exactly what the AI did, what I did, and where one ends and the other begins.
The question that remains
If the watermark does not distinguish between generated, assisted, and dictated text. If it can be removed with available tools. If public detectors do not read it. If there is no universal verifier. If it legally does not grant copyright to the AI company. If it degrades with the editing that any responsible professional does before publishing.
Then who is the watermark for? What problem does it solve that explicit editorial transparency does not solve better?
I do not have the answer. But I know that marking someone's text without distinguishing whether they wrote it, dictated it, or edited it for an hour is a problem the industry will have to solve. And until it does, the responsibility of demonstrating the authorship of your work is not in the tool's watermark. It is in the provenance chain you build yourself.