Anthropic tried to make AI-written text traceable. Within days, the internet made it untraceable again. On August 11, 2026, Anthropic announced that Claude models released since August 2 embed an invisible, machine-readable watermark into the text they generate, a way to signal this was written by AI without changing how the text looks to a human reader. The move was designed to help comply with the European Union's AI Act. Almost immediately, coders published tools to strip the watermark out, and the project went viral.
What Anthropic built, and why
The watermark is a hidden statistical pattern woven into word choices and phrasing. You cannot see it, but a detector that knows what to look for can read it and flag the text as Claude-generated. The goal is transparency of origin, part of a broader regulatory push in the EU to label AI-generated content so readers, platforms, and regulators can tell machine text from human text.
On paper it is a reasonable idea. Provenance and disclosure are becoming legal requirements, not nice-to-haves, and a watermark that survives copy-paste is more robust than a label anyone can delete with a keystroke.
The antidote arrived in days
Then reality showed up. Within days of the announcement, a developer released code to remove the watermark, and it spread fast: more than 20,000 bookmarks on X and over 100 contributors piling into the GitHub project. The removal methods are almost insultingly simple. Rewriting the text, reformatting it, or running it through another AI tool is often enough to break the hidden pattern, because those steps reshuffle exactly the word-level choices the watermark relies on.
In other words, the same fluidity that makes AI text easy to generate makes the watermark easy to erase. Anyone who wanted their Claude output to be undetectable could get there with a few minutes of effort and a free script.
For sales teams tired of cold leads, slow customer responses, and manual processes, Dapta is the ultimate tool.
Dapta is the leading platform for creating AI sales agents specifically designed to increase inbound lead conversion. Respond to your leads in less than a minute with voice AI and WhatsApp that converts.
If you want your team to sell more while AI handles the complex stuff, you have to try it.
What it means for marketing and compliance teams
If you use AI to produce content, this cat-and-mouse dynamic is now your problem, not just a research curiosity. The practical questions are immediate. Will your published text be flagged as AI-generated by a detector? Does it matter for your industry, your clients, or your legal exposure? And can you rely on any watermark to prove, or disprove, how a piece of content was made?
The uncomfortable answer is that detection will be a permanent, moving market rather than a solved problem. Every detection mechanism arrives with its own workaround, and neither side ever fully wins. Watermarks do not create trust; they just move the conversation to who has the tools to bypass them.
For teams, the sane response is to focus on things that hold up regardless of detection: disclose AI use where it is required or expected, keep humans genuinely in the loop on quality and accuracy, and treat is-this-detectable as a compliance checkbox rather than a content strategy. The watermark story is a preview of the next few years. Provenance is going to be contested constantly, and the winners will be the ones who build honest workflows instead of betting on a technical trick that a viral GitHub repo can undo in a weekend.
