You’ve probably noticed the internet filling up with synthetic text. You read an article, and something feels slightly off. The rhythm is too perfect, the examples too generic. You want a label, a warning sign—a watermark. Regulators want it, too. They demand AI companies stamp their output so we can tell what’s real.
But here’s the truth: it’s a lie. Text watermarks will never work, and the people demanding them are playing a dangerous game.
The tools we built to protect human authenticity are quietly becoming the ultimate cleanup crew for AI.
To understand why, you have to look at how language actually works. Watermarks in images or videos are relatively easy to embed—a hidden pattern in the pixels. But text is infinitely flexible. Language permits endless paraphrase. If an AI writes “The cat sat on the mat,” a watermark might tweak the statistical likelihood of certain words. But you can just ask the AI to rewrite it: “A feline rested upon a rug.” The meaning remains, the watermark is destroyed.
Any detectable statistical pattern can be perturbed. It’s trivial to remove. The European AI Act’s basis for requiring watermarks is fundamentally flawed. It assumes language is static. It isn’t.
But here is where the story takes a dark turn.
Let’s say, for the sake of argument, the AI companies actually implement these watermarks perfectly. They tag every piece of generated text. Regulators cheer. Consumers feel safe. But what happens next?
AI scrapers—those massive bots that crawl the web to train the next generation of models—now have a convenient signal. Instead of scraping everything and getting a polluted dataset full of AI-generated garbage, they can just look for the watermark and skip it.
A watermark doesn’t prove a text was written by a human; it only proves it wasn’t written by a machine too lazy to cover its tracks.
This turns a compliance tool into an unintended feedback mechanism. The watermark, designed to prove AI authorship to readers, becomes a high-value training-data filter for AI companies. They get a cleaner internet, free of their own synthetic exhaust, all funded by regulatory mandate.
Most people miss this. They think watermarks are about transparency. They’re not. Watermarks are an operational asset for the AI industry. The “failure” of watermarks to actually stay attached to text isn’t a bug; it’s a feature. The bad actors will strip them out to deceive you, while the AI companies will use them to train better models on purely human data.
If you consume or produce text online, these watermarks won’t reliably tell you what’s human. They will only reshape how AI systems train and what content gets filtered, quietly accelerating the spread of synthetic text while the regulators pat themselves on the back.
We asked for a shield to protect us from AI, and they handed us a filter to help AI train itself.
Stop trusting the watermark. It was never meant to protect you.
FAQ
Q: If watermarks are so easy to remove, why are regulators demanding them?
A: Because regulators don't understand how language models work. They want a checkbox solution for a highly complex, fluid problem.
Q: How does this affect me as a writer?
A: If you write online, your human text will be scraped to train the next generation of AI, while AI-generated text gets filtered out using the very watermarks meant to 'protect' you.
Q: So watermarks are actually a good thing?
A: For AI companies, absolutely. It's a free, automated data-cleaning pipeline funded by compliance mandates.