THOUGHT LEADERSHIP

AI watermarking isn’t the real problem. Our shame around using AI is.

Anthropic’s new Claude models now watermark the text they generate, and OpenAI signs and marks its images and audio. The loudest response has been people asking how to remove it. That reaction tells us more about where we are with AI adoption than the technology does.

AI Watermarking
Content Provenance
AI Governance
AI Strategy

19, AUG 2026 · Dr Lara Okunuga · 8 MIN

The recent debate around AI watermarking has focused heavily on one question: how do you remove it?

That reaction is revealing. Instead of asking whether we should stop using AI to generate or improve content, much of the conversation has shifted to how people can keep using AI without anyone being able to prove it. TechCrunch reported users cancelling subscriptions over the change, while others pointed out the obvious: the main reason to object is not wanting anyone to know.

The technology is interesting. The response to it tells us more about where we actually are with AI adoption.

AI has become useful enough that millions of people rely on it for writing, research, coding and editing. Businesses are actively building it into their workflows. Yet there is still something uncomfortable about attaching the words “AI-generated” to a finished piece of work.

We’re perfectly happy to use AI. We’re just less comfortable admitting how much.

We’re perfectly happy to use AI.

We’re just less comfortable admitting how much.

That tension is getting harder to ignore now that Anthropic and OpenAI are both investing seriously in watermarking and provenance.

What is AI watermarking?

Briefly, because the internet is already full of technical explainers. An AI watermark is a detectable signal left in AI-generated content, usually invisible to the person reading or viewing it.

Anthropic announced that future Claude models will generate watermarked text, using a version of Google DeepMind’s SynthID-Text approach. It subtly influences some of the low-stakes word choices the model makes, creating a statistical pattern that can be checked with a key. No hidden characters, and Anthropic says no effect on quality. New Claude models in the EU from 2 August 2026 support it, and it applies beyond Europe. OpenAI’s provenance system combines signed C2PA metadata with SynthID watermarks in supported images and audio, checkable through its public verification tool.

Two details matter for everything that follows. Detection only estimates the likelihood that AI was involved; it weakens on short or edited text, and heavy rewriting can remove it. And Anthropic is clear that the watermark cannot distinguish between Claude writing something and Claude merely editing it, and says nothing about who owns or answers for the finished work.

The interesting question is why any of this makes people so uncomfortable.

We accept AI, but still feel the need to hide it

AI itself is being normalised incredibly fast. People run their emails through ChatGPT or Claude. Developers lean on coding assistants. Marketers use it for drafts and research. We see it constantly with the SMBs we work with at FreshStack: the appetite to use AI is real, and so is the quiet worry about whether clients will “find out”.

Someone will happily say they used Grammarly to edit an article, Google to research it and Photoshop to prepare the images. Say Claude wrote the first draft, and the perception changes immediately.

The difference is that generative AI doesn’t just assist with mechanical tasks. It participates in the part of the process we associate with human thought. When software fixes a spelling mistake, nobody feels the author is threatened. When software suggests the argument and rewrites the paragraphs, the boundary blurs.

Watermarks force us to confront that ambiguity.

“AI-generated” isn’t a binary anymore

The biggest weakness in the debate is the assumption that content divides neatly into two piles: written by a human or written by AI. Real workflows don’t look like that.

One person spends hours researching a topic, develops the argument themselves, hands it to Claude for a first draft, then rewrites sections, checks every claim and publishes. Another asks a model to generate an entire article on a subject they know nothing about and publishes it unread. Both technically involve AI. They are not remotely equivalent, and there are dozens of variations in between: AI as brainstormer, AI as editor, AI challenging the author’s assumptions without writing a word of the final copy.

There is a deeper problem with the label, though. It assumes the AI is the author and the human is absent. In reality, AI doesn’t generate content in the dark. Someone decided the piece should exist, chose the angle, briefed the model, rejected the outputs that missed, and shaped what survived. That is a director’s role. Nobody describes a film as “camera-generated” because the director didn’t physically operate the equipment. The vision, the taste and the judgement sit with the person directing, and credit follows the direction, not the tool.

A watermark can’t see any of that. It marks the tool that produced the words, not the person whose ideas the words carry. Which raises an uncomfortable question for anyone anxious about the “AI-generated” label: how much of our own thinking are we quietly handing credit for to the software that typed it up?

Directing cuts both ways, of course. If you claim the vision, you own the result.

Responsibility matters more than removal

Publishing AI output without checking it is a problem. Flooding search with thousands of automated junk pages is a problem. Fabricating images or political material to deceive people is a serious problem. But none of these are created by the mere presence of AI. Humans wrote inaccurate content, produced spam and misled audiences long before generative models. AI changes the scale and speed. The real issue is still accountability.

The goal of provenance should not be proving a human typed every word. It should be establishing whether someone is willing to stand behind the result. A person who publishes AI-assisted work answers for its claims and conclusions, which means actually reviewing and testing the output rather than trusting it blindly. (It’s the same discipline behind why one good output doesn’t prove a system works.)

Even OpenAI draws this line: detecting a signal indicates content likely came from its tools, but says nothing about whether it’s accurate or honest, and no signal doesn’t prove something wasn’t AI-made.

A watermark can tell us something about origin. It cannot tell us whether the content deserves to be trusted.

A watermark can tell us something about origin.

It cannot tell us whether the content deserves to be trusted.

Maybe provenance shouldn’t be embarrassing

There are contexts where disclosure genuinely matters: journalism, academic research, legal documents, politics. Nobody should pretend every use of AI deserves identical treatment.

For ordinary business and creative work, though, we may need to become far less precious about it. Calculators didn’t eliminate mathematicians. Photoshop didn’t eliminate designers. Search engines didn’t make researchers obsolete. Each tool changed how the work was done, then became an ordinary part of it. Generative AI cuts deeper into knowledge work, but the principle holds: the question is less whether someone used a tool, and more what judgement they brought to the work produced with it.

In that world, provenance becomes useful without being shameful. Content can transparently show that Claude or ChatGPT contributed without implying the person publishing it did nothing. That would be healthier than where we’re drifting now, where people enthusiastically adopt AI behind closed doors while searching for ways to remove the evidence.

The real debate is about trust

Anthropic is introducing its watermark partly to comply with the EU AI Act’s transparency rules (we broke down what EU AI Act enforcement looks like in 2026). The infrastructure will never be perfect: watermarks can be weakened, metadata stripped, and OpenAI itself says no single technique is enough on its own.

But the bigger challenge isn’t technical. We’re entering a world where AI is deeply embedded in how people work, while culturally we still pretend that using it diminishes the finished product. That contradiction can’t last.

The question will eventually stop being “did AI create this?” The more useful question is “who directed it, and who stands behind it?”

The strangest thing about the current debate is that we’re building technology to make AI involvement more transparent at exactly the moment people are hunting for ways to hide it.

Sources

We’ve accepted AI as a tool. We just haven’t accepted what it means to admit we’re using it.

Let’s talk

Free · 30 minutes · No pitch