The watermark paradox: Are we protecting human authorship or AI?

By Marie-Celine Merret Wirstrom, Co-Founder, MC&V

Anthropic has begun introducing invisible watermarks into Claude-generated text, while other major AI companies are developing similar provenance infrastructure across text, image, audio and video. The immediate reason is regulation. The EU AI Act now requires generative AI providers to make synthetic content machine-readable and detectable as artificially generated.

But regulation only explains why this infrastructure is being implemented now. There may be a bigger reason why distinguishing human from synthetic content is becoming so consequential. The first generations of large AI models were trained on an internet overwhelmingly created by humans. Future models won't have that same internet. AI-generated content is rapidly becoming part of the information ecosystem future models will learn from, meaning models increasingly encounter the outputs of other models — potentially even descendants of their own.

Research published in Nature has demonstrated the risk of "model collapse": when models are recursively trained on model-generated data, errors can compound, diversity can diminish, and their representation of the original human data distribution can deteriorate. So watermarking isn't only useful for telling us what was made by AI. It's also useful for telling AI companies what was made by AI, allowing them to identify or manage synthetic material when assembling future training datasets.

But is that really provenance?

We're building increasingly sophisticated provenance for what comes out of AI without equivalent provenance for what went into it. A watermark might tell me an image came from Google's model. But as synthetic media becomes an ordinary part of what we consume, knowing which model produced something matters less.

The more interesting questions sit on either side of the model: What human-created material helped build the model capable of producing this output? And how much of the resulting work was actually created by the human using it? A watermark doesn't adequately answer either. It takes us back to the model and stops there. It doesn't take us through the model to its human creative history, nor can it tell us whether the human contribution to an output was 5% or 95%.

We risk creating the appearance of provenance without getting to the bottom of provenance.

For advertising, that raises another question: who is provenance actually for? Inside the industry, it matters enormously. Brands and agencies need to know whether talent consented, assets were licensed, models were approved and rights are defensible.

But does the consumer need the provenance of every tool? Advertising has always manufactured reality. We retouch photography, build impossible environments in CGI, composite performances and animate products. We've never labelled a piece of work because we used Houdini.

Coca-Cola, for example, chose to disclose AI use in its Christmas campaign with an on-screen "Created by Real Magic AI" message. But did that information help consumers understand the work, or simply change how they judged it? The campaign became a lightning rod for criticism precisely because it was AI-generated, even though Coca-Cola reported that consumer testing scored the work highly.

As of July 2026, Google Ads itself introduced a "How this ad was made" disclosure which is quite amusing when their first AI-generated ad back in October 2025 ", Planning a quick getaway" with a turkey as the hero, all made in Veo 3 didn't have any labels initially because they claimed that consumers don't care whether an ad is made with gen AI or not.

Perhaps provenance matters when it changes what consumers are being asked to believe. Was that really the celebrity? Is that a real human giving a testimonial? Is the product actually capable of what I'm seeing? That requires a more sophisticated definition of provenance than simply "AI was used".

And then there's authorship

If I write an article and ask Claude to tighten three paragraphs, what exactly does its watermark authenticate? That Claude touched it? Wrote it? That the article is "AI-generated"?

The same problem exists in production. One filmmaker might conceive, write and direct a piece, shoot real performers and use AI for environments. Another might generate an entire piece from a prompt. Both could carry an AI provenance signal. In our commercial work, the boundaries are even messier; AI-generated imagery is composited and retouched by artists. Motion graphics still involve manual animation, and hybrid productions combine real talent, generative environments, traditional VFX, and post-production, where much of the workflow is still the same. How do you meaningfully label that?

A binary of watermarked = AI, unwatermarked = human - bears little resemblance to how creative work is actually being made.

And then we return to the systemic issue; Human culture trains AI on AI generated content. That content enters the internet, and the internet becomes training data for the next generation of AI. As that loop accelerates, knowing what is genuinely human-created may become increasingly valuable not just to us, but to the AI companies building the next generation of models.

That's the paradox. We're getting much better at marking the content humans create with AI, while still having remarkably limited visibility into the human content used to create the AI itself.

So perhaps the important question isn't whether we'll finally know when something was made with AI. Increasingly, that's old news. Are we building provenance to protect human authorship, or are we building provenance to protect AI from the consequences of an internet increasingly filled with AI