AI-Generated Content Is About To Be Traceable. Sort Of.
Anthropic will soon watermark Claude’s text to comply with EU rules. The detector everyone’s picturing doesn’t exist yet.
On Friday, August 14, Anthropic, the builder of Claude, announced it would do something that for years many AI companies have avoided: It will soon, for the masses, make its own writing easier to spot. Claude’s text will, in the near future, carry a hidden signature, buried in the model’s word choices, that Anthropic can check against a key only it holds. The EU AI Act required the move. Time will tell if more countries will adopt something similar.
Somewhere, an AI marketing director just felt a chill down the spine. (No, not me, and I’ll explain that later.)
But hold that thought, because the doomsday detector everyone is picturing doesn’t quite exist yet, and when it arrives it will prove less than you’d assume.
How Watermarking Works
When prompted, large language models or LLMs like Claude write one word at a time. At each step, the model weighs a set of choices that would all fit the sentence, and many of those choices carry no real consequence for meaning. Whether a sentence about weather describes the sky as “overcast” or “gray” changes nothing for the reader. Under normal operation, a random number settles the tie and fills in the word.
The model constantly hits forks where two words work equally well. Normally it picks by something like a coin flip, and the result is arbitrary. Watermarking replaces that coin flip with a rule. The rule takes the words already written, runs them through a secret key, and produces the choice. Nothing changes for you, the end user. The output reads the same, and the sequence still looks arbitrary. But Anthropic, which holds the only key, can run a passage back through the rule and calculate the odds that Claude produced it.
Anthropic reports no impact on content, creativity, or readability in internal testing, and cites research from Google DeepMind, published in the journal Nature in 2024, in which human raters comparing watermarked and unwatermarked answers side by side found no difference in quality. Most readers and editors won’t notice. (Unless you’re droning on about “in a world” or on a “landscape” or the “not just this, but that” construction.)
This AI stuff is heady stuff, so I enlisted Claude here to distill the jargon-y things and to build me this handy dandy infographic on what its watermark will and won’t do.
It is worth noting what doesn’t exist yet: the detector. Anthropic says a tool is coming and hasn’t set a date. When it lands, you still won’t hold the key. You’ll send text to Anthropic and get a probability back.
Two things change here, and neither is about catching writers red-handed. You may soon have to answer a question nobody has ever asked you: How was this made, and who checked it before it went out? Europe now treats that as worth answering, and for the first time there’s a way to partially verify the answer. If you sell into healthcare, finance, or anything that runs through procurement, expect it in an RFP within the year.
The most uncomfortable version of this question isn’t for the agency that puts out AI-generated content. It’s for the client who never thought to ask. Plenty of companies are paying premium rates for content they assume was written the way the invoice implies, and they have no way to confirm it. That was a fine assumption last year. It’s a weaker one now that the question has an answer.
So let me answer it about us first.
Yes, We Use AI. Here’s What That Actually Means.
Earlier I said no chill went down my spine. Here’s why.
Triple7Marketing uses AI. Our writers use it for research, for outlines, for the second and third pass at a stubborn paragraph. Anything that runs through those tools will carry whatever watermark the provider applies, ours included. Any agency telling you otherwise in 2026 is either behind, or counting on you not to ask.
Search engines and AI systems are both hunting for the same thing: evidence a qualified human stood behind the work. You can’t fake that signal. You can only pay for it.
Every Triple7Marketing campaign passes through named writers and editors who interview your team, check statistics against the original source instead of somebody’s summary of it, and cut the sentences that sound authoritative while saying nothing.
That editorial layer is what produces the experience and expertise signals that both readers and search systems reward, and it is the reason our clients’ content survives contact with skeptical audiences.
A watermark cannot tell you whether a piece of writing is accurate, original, or worth reading. Editorial process can.
If you want to know what AI systems currently say about your brand, and where your content leaves you exposed, that’s what our AI Visibility Audit is for.
So, Claude, how did we do?
Sheena Foster is the founder and CEO of Triple7Marketing, a digital marketing agency specializing in SEO, GEO, and earned, paid, and owned media. Before founding it she spent 23 years in journalism and communications, reporting on CNN, ABC, and NBC, with bylines in Essence, Reader’s Digest, and TheGrio, and holding final editorial sign-off across nearly 200 clients including Fortune 500 companies and executives in high-stakes reputational moments. Schedule a free strategy call with the Triple7Marketing team.