Daybreak - Everyone wants AI content labelled, until the label lands on us

Episode Date: August 12, 2026

Earlier this week, Anthropic began stamping an invisible watermark into the text and images Claude generates. The move came in order to comply with the EU AI Act but it will be switched on worldwide. It is the thing teachers, editors, and hiring managers had been demanding for years — a way to finally tell human work from machine. Within hours, many people who pay for Claude were in revolt, calling it a scarlet letter on their own work. But the mark barely does what either side thinks, right now. And the people it exposes are rarely the ones gaming the system. So why do we keep asking to know what's AI, only to step back when the answer might be us?Tune inDaybreak is produced from the newsroom of The Ken, India’s first subscriber-only business news platform. Subscribe for more exclusive, deeply-reported, and analytical business stories.

On Tuesday, Anthropic gave a lot of people something close to what they'd been asking for, and within hours, many of the same people were furious about it. The company said that its AI assistant, Claude, will start stamping an invisible mark into the text that it writes and the images that it makes. You would never really see it. It is there so that later, in theory, somebody could check whether a piece of work came from the machine. For years now, teachers, editors and hiring managers have been pleading for exactly this, a way to tell some kind of a signal that separates what a person made from what a chatbot produced.

But now that the signal has arrived, the loudest reaction from the people who use these tools is not really relief. It is rage. Paying customers said that it is like being branded with a scarlet letter. They said that they would switch to another open source tool. On the biggest clod forum on Reddit, the verdict after nearly 400 comments was close to unanimous and it was not kind. So here is what is strange about this whole thing. We say that we want to know what is air generated. We talk about it all the time.

But when a company builds the thing that tells us, we recoil from it. So you see, there is this gap between the transparency that we demand and the transparency that we can handle or actually stop. Claude's watermark just happens to be the thing that exposed it. Because the closer you look at this fight, the clearer it gets that this is not a simple case of a company getting it wrong, though it may have. It is a genuinely hard problem. And the watermark itself, the thing that everybody is arguing about, does far less than either side seems to think.

Welcome to Daybreak, a business podcast from the Ken. I'm your host, Nick Da Sharma, and I don't choose the news site. Instead, every day of the week, my colleague Rachel Verghese and I will come to you with one business story that's worth understanding and worth your time. Today is Thursday, the 13th of August. Let's start with what this watermark actually is, because you cannot hate or dislike something that you cannot picture. Now, if you're imagining it as some kind of a hidden character or a digital version of a tag

stapled into the file, that is not really it. The watermark is a kind of a statistical fingerprint pressed into the words themselves. As Anthropic says, it leans very slightly towards certain word choices over others, which are equally good ones following a secret pattern. You will notice nothing when you read the text, but if you run it through a detector built to find that pattern and across a long enough passage, the fingerprint will surface. Researchers call this general method green listing. Anthropic has not said exactly how its own version works. Images are handled in a different way with a signed label attached to each file rather than woven into it.

Also, turns out a limit is baked into the text method because Claude is talking to you live word by word. So it stamps the watermark on the fly which caps how deep it can go. Now, the real question, why would someone with nothing to hide still want this gone? Actually, it has very little to do with cheating. It is about a kind of a penalty that has settled into how we work. A writer on Medium, Karen Isabel Noop, described it as a sort of a purity culture of forming around AI. Like, if you say this was written without AI, it is a badge of owner. And if you say AI assisted, it is like a stigma.

Work is splitting into clean and processed. And here is her sharper and more important point. She says a lot of people did not choose AI freely. They were pushed, sometimes even pressured into using it. But now, the same people are shamed for having done it. The thing is, this judgment already existed and I will not be getting into the reasons why today. But what this watermark basically does is hands everyone a switch to turn it on. As one Claude user put it, the watermark does not care whether Claude wrote the whole thing or just helped.

and everyone who sees it will draw their own conclusions. The people this hurts the most are the ones using it honestly and as a collaborator. And if you pause for a second and really think about it, this penalty would fall the hardest in a place like India. Across much of our academic and working life, English is a second or third language. And a huge number of people reach for these tools to make their writing read cleanly. They were already the ones getting wrongly accused because the old

detectors kept flagging plainer and non-native English as machine-written. And now, the tool that they would use to fix that leaves a watermark that invites the accusation for real. For a developer, it is the code base, where the mark clings less to the code than to the commit messages and pull request notes wrapped around it. And in some circles, that tag reads as an insult. There is an honest argument on the other side as well. Some people think that the discomfort is the whole point.

If a machine did the work, hiding it is dishonest and a watermark just makes the truth visible. Fair enough. But here is where it gets strange. The tag that people are so desperate to avoid may actually barely be doing the job that they fear. More on that in the next segment. The thing is that the person who actually wants to pass AI work off as their own can get rid of this watermark in minutes. Here is what an AI detection company GPT Zero's Alex Quay, whose own company builds AI detectors, said. The pattern does not hold up to heavy paraphrasing. The free paraphrasing tools that

he tried quickly slipped past Google's version of the same technology that Claude is using. And the research agrees. A team from Maryland and Howard ran AI text through a rephraser a few times and watch the detection collapse from almost perfect to almost nothing, while the writing still read okay. So, you don't even need a clever trick. Actually, Anthropics says so itself. On its own help page, the company lists what makes the watermark weak. Heavy editing, paraphrasing, translation, mixing the text into other writing.

As an anonymous Anthropic engineer put it quite plainly, this watermark is not perfect. The image version is even weaker because the label that sits on the file falls off the moment someone resaves it or grabs a screenshot. And then comes the trap. To let people check for this watermark, Anthropic has to hand out its own detector. But the moment that it does, as Gwe wants, anybody can run their text through it again and again editing until the watermark is gone. But if Anthropic keeps the detector private instead, only it can verify if content is AI generated or not. So there is basically no clean way out of this box.

At this point, you might ask, why bother at all? And here is where the even harder part of this dilemma lies. Knowing where the content came from genuinely matters. Forget about the student. Think about the woman whose face is stitched onto a video that she never made. or the voter who answers a phone call in a politician's cloned voice. For them to be able to prove that the thing was faked is their protection. And that is a real case and it is why this problem is not going to go away.

The push comes from Europe's new AI law, which nearly every major lab has signed including OpenAI, Google, Meta and Microsoft. But this watermark just covers texts and images right now. The things that actually frighten us about, AI, it doesn't touch. If you have any thoughts on this subject, I would really like to hear them write to me at Snigda, S-N-I-G-D-H-A, at the ken.com. That's the-hyphen-ken.com.

Or even better, share this episode on social media with your thoughts or questions and start a conversation about this. Do not forget to that again.