The Real Watermark is You

I don’t care about Claude’s watermark.

That’s probably not the opening statement you expected from someone writing about Anthropic, Claude, and the growing controversy around AI-generated text. I really don’t.

Watermarks are nothing new – we’ve been putting them on images for years, along with embedding information in audio, and altering pixels in ways undetectable by the human eye but instantly detected by machines. Attaching cryptographically-signed metadata to digital content is nothing new, and now we’re doing the same things with text.

The technology used is far more interesting than frightening – and I am fascinated by why so many people are suddenly terrified by it. I think the sense of terror is coming from a lack of confidence more than any other factor.

Before I get accused of writing an anti-AI piece: I’m sitting here at a desk with multiple AI agents running, a dedicated inference machine, automation systems, and enough hardware to make the average “AI will replace your entire staff” LinkedIn post look positively quaint. I use these tools constantly – I just don’t let them do the thinking for me.

So, What is Anthropic Doing?

Anthropic has started implementing an imperceptible statistical watermark in newer Claude models. The important part here is statistical. There isn’t a hidden Unicode character sitting somewhere in your document.

The basic idea behind statistical text watermarking is that when a language model has several reasonable tokens it could choose next, it can subtly bias those choices according to a predetermined pattern. A single choice or sentence doesn’t necessarily mean anything, but across a large body of generated text, those choices can create a statistical signature that a detector can potentially identify.

That means the watermark isn’t something that can be pointed out, but something that can be measured instead.

This isn’t a technique Anthropic invented because it decided to ruin everyone’s LinkedIn posts one morning. Statistical watermarking is a much larger body of research into AI-generated content provenance. Other companies have been working on watermarking for images, audio, video, and text for years.

Google’s SynthID embeds detectable signals into supported generated content while OpenAI uses C2PA Content Credentials and SynthID for supported media. The European Union’s AI Act is also pushing the industry toward machine-readable ID and provenance mechanisms for AI-generated content. While the technologies vary, the goal is more or less the same – making it possible to establish something about where digital content came from.

There’s an important distinction to be made here. A watermark can provide evidence of provenance, but it doesn’t magically establish authorship. If I write 1,500 words myself and use an AI system to proofread them, the resulting document may contain evidence that an AI system was involved. The evidence does not mean the AI wrote the 1,500 words.

Likewise, if a document contains no detectable watermark, that doesn’t prove a human wrote every word. Provenance is not authorship; they’re two different things, and I suspect we’re going to need to get considerably better at explaining that distinction.

We’ve Been Doing this Forever

One reason I find some of the reactions to AI watermarking amusing is that the underlying idea isn’t exotic. We’ve been embedding information into digital media for a very long time.

Take images, for example. A watermark can be embedded into the pixel data in ways that are invisible to the person looking at the image. The image appears completely normal, but tiny changes in the underlying data can encode information that an algorithm can recover. Depending on how it’s implemented, the signal may survive things like compression or resizing – or it may not.

Audio works similarly. An audio signal can be represented as a spectrogram, essentially showing frequency over time. Information can be embedded into the signal in ways that aren’t obvious when you listen to it – but can be detected computationally. Again, nothing strange going on – the information is simply there.

Looking at cryptographic metadata and systems like C2PA,provenance information is attached to the content rather than hidden inside the content itself. Think of that less like invisible ink, and more like a signed chain of custody.

Now we’re applying related ideas to generated text, leading to the  question: how do we establish something about the history of a piece of digital content?

When somebody discovers that Claude can produce text with an imperceptible statistical signature and reacts as though we’ve entered some unprecedented technological nightmare, my first response is, “Take a deep breath, this is nothing new.”

Why Are People So Upset?

There are very legitimate reasons to be concerned about AI-generated content being identifiable. People worry about employment, academic work, client relationships, publishing, professional reputation, and automated detection systems making incorrect accusations.

Those are real concerns – but there’s another group that feels burned by the AI industry more broadly. They look at how AI companies acquired enormous quantities of human-created material from the web and build enormously capable models from it, then watch those models generate content that compete with the people whose work helped make the models useful. That screams “bait-and-switch.”

There are legitimate copyright, compensation, and consent questions here. The legal battles surrounding AI training data aren’t going away simply because the models are useful.

I also think there’s another group reacting to the watermarking story for a completely different reason – they’re afraid someone is going to find out they used AI, and that got me thinking about confidence. Not confidence in the AI systems being used, but confidence in oneself.

Everyone Else is Doing it

Spend enough time on LinkedIn, maybe 15-30 seconds, and you’ll see a remarkable transformation in the way people talk about AI. There are stories about people sitting around drinking iced coffee while their AI agents run their business, along with stories about companies that are supposedly almost completely autonomous. Stories of people waking up late, checking a dashboard and discovering their collection of AI employees made money while they slept are undoubtedly true, and that’s what makes them effective.

What you don’t see are the 9,997 people who built an “autonomous business” that made them $38.46 – you see the person who made $380,000. Instead of reading about the person who spent six months automating something that saves them four hours a week, you see the person who claims they replaced an entire department.

There’s an incentive to storytelling – the bigger the claim, the more attention it gets for the person who posted it. Eventually, somebody looks at all of this and thinks, “everyone else is using AI like this – Am I falling behind?”

That’s a confidence problem with a technology wrapper, and something we’ve seen countless times before. If the person to your left and the person to your right are using generative AI, shouldn’t you be using it too? If not, they have a competitive advantage, right? Eventually everybody uses the same approach and starts sounding the same.

Knowing What You Want

I use generative AI extensively – and not because  I don’t trust myself, but because I know what I want from it. When I’m writing something substantial, I’ll often develop a skeleton of what I want to say first. Sometimes, I’ll use an LLM to help me build the skeleton – I’ll ask it to identify things I haven’t considered or detect points where bias is at play. Based on what I get back, and what’s developing in my head, I may have it generate a rough draft that I can work against – and then I sit down and write.

I have a “map” on one side of my screen and a blank space for my actual writing on the other, and that’s fundamentally no different from how I’ve worked for years. Back when I worked at Valiant, I created templates for people who needed to write blog posts. The templates helped them understand what a useful piece of content needed to contain – an introduction, context, the actual point with supporting information, and a conclusion that doesn’t wander off into the technical words.

The major difference today is that an LLM can help me create the skeleton much faster. The writing, judgment, and responsibility are still mine, and so is the voice.

A Lesson I learned 25 Years Ago

I attended Pace University around the turn of the century. One of the professors who had a lasting impact on me was Dr. Mark Hussey, a Distinguished Professor Emeritus of English who taught literature for more than four decades before retiring in 2021.

I remember him for a lot of reasons, and I also remember that whenever he walked into class, my voice was usually the first he heard.

“Dr Hussey! How’s the wife and kids?”

He’d laugh, we’d have a brief conversation, and then class would begin.

I was oddly vocal in a lot of my classes – part of that is just who I am, and part of it was a deliberate “enhancement” of myself. From the moment I walked into the main building for orientation, the place felt strangely comfortable. There were hundreds of students, and I didn’t particularly want to disappear into the sea of them. I didn’t. Instead, I talked, joked, and asked questions; I differentiated myself.

I didn’t realize at the time how much of that would become foundational to how I wrote. The first paper I submitted to Dr. Hussey was dry and factual, reading more like a reference manual. This made sense, considering most of what I was reading at the time. We talked about it, and he encouraged me to bring more of myself into the writing. My tone, sense of humor when appropriate, and my tendency to wander down a tangent before eventually finding my way back to the point.

Dr. Hussey essentially told me, “That’s you. That’s what needs to be on the paper.” Those aren’t his exact words; it’s been over two decades, but they’re close enough that I remember the lesson. I took his advice and the next paper I wrote was used as an example for the class. That didn’t mean it was perfect by any means – it meant that it was me, and a distinction that has stayed with me for more than a quarter of a century.

Red Queen, Light Edition

I’ve written an enormous amount of material since then – analysis, documentation, educational materials and training content, marketing copy, blog posts, technical explanations, scripts, and so on. There were ideas that should have worked, but didn’t – along with ones that lead to outcomes much better than expected.

All the work has been mine, and generative AI hasn’t changed that – it has, however, changed how I work, and that’s not the same thing. I don’t think the answer is to reject AI and pretend that using it somehow invalidates the work. That’s silly. I also don’t think the answer is to dump an AI-generated block of text into a CMS, slap your name on it, and call yourself a writer. That’s lazy.

Not only is it lazy, but it also creates a particularly ugly flywheel.

Your competitor uses AI, so you use AI because your competitor uses it. Their other competitor uses AI because you do, and everyone produces more content. Nobody wants to be the one who doesn’t produce more content, and suddenly the Internet is drowning in perfectly legitimate-looking garbage. That’s the distinction I care about. I’m not talking about automated content forms; I’m talking about content produced by legitimate businesses and professionals that technically meets every requirement while saying absolutely nothing that couldn’t have been said by the next person in line. It isn’t necessarily fraudulent, but it is soulless.

Just the Facts - Thanks

There’s another reason I find the watermark controversy interesting- one of the things I use LLMs for most frequently has nothing to do with generating content: fact checking.

My ability to use an LLM to fact-check is particularly important when I’m producing an edition of Blumira Briefings. The team covers current IT and security stories, and there are talented writers involved in producing the source material. Security-related news, however, has a short half-life; after a story is published, the information changes. A vendor issues a statement, a researcher publishes additional data, the scope of an incident changes, or a technical detail gets corrected. Meanwhile, social media is still circulating the original headline.

When I’m reporting on a security incident, one of my first steps is often to give the information to an LLM and ask it to challenge what I’m reading. Was the event in the story actually confirmed? Has anything changed since the story was published? Is that CVE really being exploited in the wild, or is speculation for the sake of impressions at play?

I’m not asking the machine to tell me what happened and then blindly repeating the answer – I’m asking it to help me find the places where I might be wrong, and that’s a completely different relationship with AI.

Frankly, I think it’s one of the most valuable applications of AI. If you’re producing content that other people will use to make decisions, being confidently wrong is dangerous. The same idea applies to IT, marketing, and pretty much everywhere else.

Human in the Loop

We’ve reached a point where “human in the loop” has become almost meaningless, at least in the scope of what I’m writing about here. If the human’s job is to click “approve” after the machine finishes everything, that’s not really human-in-the-loop decision-making - it’s just an empty form of quality control. A human should be establishing the question, evaluating the evidence, challenging assumptions, identifying gaps, and deciding what matters – and an LLM can participate in all of those things.

Using an LLM can make me faster, more thorough, point out something I missed and argue with me. It can even be spectacularly wrong and force me to go look something up – which is equally useful. The machine doesn’t have to be perfect to be valuable; I just need to remain responsible for the result.

What Balance Looks Like

I have multiple agents running beside me, local inference, automation, and tools doing things in parallel that would have required considerably more manual effort just a few years ago. I use them a lot, but I’m not sitting here “putting the machines to work for me” while I go drink iced coffee.

Not only that, but I’m still here at my desk – thinking, writing, questioning, and making decisions for myself. The machines are leverage, not a replacement for thought. I think that’s where the balance we’re going to have to figure out as the systems become more capable – use AI when it makes you better or faster. Incorporate it into your efforts to challenge your assumptions and fact-check your work. Use it to organize ideas in your head to find holes in our argument before you begin writing - but don’t use somebody else’s AI-powered lifestyle as a measuring stick for your own worth.

Don’t assume that because someone on LinkedIn says their agents are running a $10 million business while they sleep that you’re somehow failing because you’re still doing work – and don’t assume that because Claude can tell someone that AI touched their document that the machine somehow owns the words. Most importantly, don’t outsource the part of the process that makes the result yours.

The Real Watermark is You

Maybe the conversation about AI-generated content needs a little less obsession with whether a machine can detect the machine.

Maybe the better question is what happens after the detection question has been answered. Can you explain what you wrote, and can you defend the conclusions? Can you identify the assumptions or tell when something doesn’t sound right? Most importantly, can you put the tools down and do the work yourself when that’s what the situation requires? Those things matter considerably more to me than whether a statistical watermark exists somewhere underneath the words.

Dr. Hussey taught me something about writing that I didn’t fully appreciate at the time. A piece doesn’t have to be perfect to be worth reading; it has to belong to somebody. Over twenty-five years later, I’m still trying to make sure mine does, and I don’t care whether AI helped me get there. I care that I got there, and that my output has legitimate value.

Use the machines, just don’t let them convince you that you have nothing worth saying without them. Get back to being yourself.