Since the EU started requiring AI companies to mark AI-generated content in August 2026, “AI watermarking” has become one of the most misunderstood topics in tech. Depending on which post you read, it’s either a secret surveillance code hidden in your writing, or the silver bullet that will finally let us detect AI content reliably. It’s neither. Here’s what AI text watermarking actually is, what it genuinely can and can’t do, and the myths worth clearing up – because getting this wrong has real consequences for how much you trust what you read.
First, What Is AI Text Watermarking?
AI models like Claude generate text one word at a time. At each step, the model picks the next word from a list of plausible candidates. Take “The weather today was cold and…” – the next word is very unlikely to be “sugary,” but “overcast” and “grey” are both perfectly reasonable, and it usually makes no real difference which one gets chosen. Normally that choice is settled by a random number.
Watermarking uses those low-stakes choices to leave a subtle statistical pattern across a piece of text. Instead of an arbitrary random number deciding between “overcast” and “grey,” the choice is steered by a secret key combined with the preceding words. The words are still effectively random and the meaning is unchanged – but if you hold the key, you can later check whether a passage’s word choices match the pattern the key would produce, and assign a probability that the text came from that model. Anthropic’s approach is based on Google DeepMind’s SynthID-Text method, published in a peer-reviewed Nature paper, and it’s rolling out because the EU AI Act now requires major providers serving the EU market to mark AI-generated content.
With the basics established, let’s go through the myths – the overblown fears on one side and the overblown hopes on the other. If you want the broader grounding on how these models work in the first place, our plain-English explanation of what AI is covers it.
Myth 1: “They’re Hiding Secret Characters or a Tracking Code in My Text”
Reality: nothing is added to the text, and the watermark carries no identifying information. This is the “genius and a little creepy” fear doing the rounds, and the creepy half is largely misplaced. There are no hidden characters, no invisible Unicode, no metadata payload embedded in the words. The watermark isn’t a thing inserted into the text; it’s a statistical pattern in choices the model was already making at random. Crucially, it carries no information about who generated the text – it can’t be traced to a specific person, organisation, or individual chat. It doesn’t add tokens and doesn’t cost more. It answers one narrow question – “was this likely generated by this model?” – and nothing about who did it or why.
Myth 2: “Watermarking Degrades the Quality of the Writing”
Reality: there’s no measurable impact on quality, creativity, or readability. Because the watermark only influences genuinely low-stakes choices – the “overcast versus grey” decisions where either word is fine – it doesn’t push the model toward worse or weirder words. It won’t make the model reach for some obscure synonym it would never normally use. When DeepMind tested this by serving watermarked responses to a share of real Gemini traffic and comparing user thumbs-up and thumbs-down ratings, they found no statistically significant difference from the unwatermarked model. In side-by-side human comparisons, raters saw no quality difference. To a reader, watermarked and unwatermarked text are indistinguishable – which, incidentally, is the opposite of a banknote watermark that you can hold up to the light.
Myth 3: “Now We Can Finally Detect AI Content Reliably”
Reality: this is the big one, and the honest answer is no – not even close. Watermarking is far more fragile than the headlines suggest, and researchers are openly sceptical it can curb the flood of “AI slop.” The limitations are fundamental, not teething problems:
- Rewriting destroys it. Google itself admits SynthID’s detection accuracy is “greatly reduced” when users thoroughly rewrite generated text, and that it “is not designed to directly stop motivated adversaries.” Paraphrasing tools – swapping words for synonyms – can remove a text watermark almost entirely.
- Free tools can strip it. A TIME journalist with no engineering background reported using a free online tool to remove a SynthID watermark, and stripping image metadata simply by taking a screenshot.
- Some text can’t be watermarked at all. Low-entropy content – factual answers, code, anything where there’s really only one correct way to say it – has no “low-stakes choices” to hide a pattern in. There’s nothing to watermark.
- Open-weight and non-compliant models don’t carry it. Anyone using a model that doesn’t implement watermarking produces text with no mark at all. And there’s no shared cross-provider standard, so a checker has to run many different detectors while an evader only has to defeat one.
In short: watermarking can offer a probabilistic signal about unaltered text from a cooperating model. It cannot reliably catch someone who is actively trying not to be caught. That’s a useful distinction to hold onto.
Myth 4: “If Text Isn’t Watermarked, It Must Be Human”
Reality: this is the most dangerous myth of all, and it’s the inverse of the last one. If watermarks can be stripped, if some AI text is never marked, and if human-written text obviously has no AI watermark either, then the absence of a watermark tells you almost nothing. Yet the natural human instinct is to treat “no watermark detected” as proof of authenticity – and that false confidence is arguably worse than having no system at all.
We’ve been here before. In the early web, a slick, well-designed website signalled a credible institution, and a janky one screamed “scam.” Then cheap tools made professional-looking sites available to anyone, and that visual signal became worthless – yet people kept trusting it. The same trap is forming around AI detection: treating a clean detector result as a green light. The safest mental model is that a watermark’s presence is weak positive evidence, while its absence is not evidence of anything.
Myth 5: “This Is a Claude / Anthropic Surveillance Move”
Reality: it’s a regulatory compliance requirement across the whole industry. As of 2 August 2026, the EU AI Act requires providers serving its market to mark AI-generated content. Anthropic is implementing watermarking for future Claude models to comply, but so are the other major developers – OpenAI and Google among them have signed the same Code of Practice and are rolling out their own approaches. This isn’t one company quietly tagging its users; it’s a legal transparency obligation being met, imperfectly, across the sector.
What This Actually Means for You and Your Business
Strip away the myths and here’s the practical takeaway for anyone creating or relying on content:
- Don’t build policy around AI detectors. If you’re a business, a publisher, or an educator tempted to police content with an AI detector, understand that they can be fooled by simple rewriting and produce false results in both directions. Making consequential decisions – rejecting work, accusing someone – on a detector’s say-so is a genuine risk.
- Don’t treat “not AI” as a mark of quality or trust. Whether content is watermarked matters far less than whether it’s accurate, useful, and genuinely helpful. That’s the standard that holds up regardless of how it was produced – and, not coincidentally, the same standard that search engines and AI systems increasingly reward.
- Value provenance signals for what they are. Watermarking is a reasonable transparency tool for identifying unaltered AI output from cooperating providers. That has real uses – flagging AI content at scale on a platform, for instance. Just don’t mistake a transparency signal for an enforcement mechanism.
- Keep humans in the loop. The reliable way to know whether content is trustworthy hasn’t changed: verify claims against sources, check the reasoning, and judge the substance. No watermark replaces that.
The Bottom Line
AI text watermarking is cleverer and less sinister than the scare stories, and far weaker than the hype. It doesn’t harm your writing, doesn’t hide a tracking code, and doesn’t identify you. It also won’t reliably catch determined bad actors, can’t mark every kind of text, and – most importantly – can never make the absence of a watermark mean “this is genuine.” Treat it as a modest transparency measure, not a truth detector. The moment we start trusting content because a checker stayed quiet is the moment watermarking makes us less safe, not more.
If you want help thinking through how AI content, detection, and transparency affect your own business – content strategy, policy, or just separating the real risks from the noise – get in touch.