Anthropic on Monday confirmed that it embeds an invisible watermark in text generated by Claude. It applies to models released since August 2 and holds worldwide, and for files the company used the C2PA standard. The mark survives copy and paste and, according to Anthropic, some edits as well. But the most interesting sentence in the whole announcement is this one: the mark shows that Claude took part in the text, not that it wrote it. Even a requested proofread of a single paragraph leaves a trace.
In my view, that settles the dispute over labeling AI content before it has properly begun. A watermark is a touch detector. And today, with professional writing, no technique can tell touch from authorship, because co-authorship isn’t an on/off state but a spectrum with ten steps. I should note that the European Commission apparently means it this way and claims nothing more: the goal is to flag that AI took part in a text, not to determine who the author is. But the reader will read the label as a verdict on authorship, and there is only one way to deal with that: show them what the collaboration looks like.
On Vibecoding.cz I have a footer stating that the content is co-created by artificial intelligence. So let’s look at what that sentence actually means.
How one article comes about
Right now I’m writing a piece about Meta’s new models. Hypata prepares most of the source material for me. It goes through my older articles, pulls out what I wrote about those models earlier, and adds what has changed since, including the renaming of the Llama line to Muse and Yann LeCun’s departure.
Hypata is my layer on top of Claude with access to my archives. Anyone else can use the bare model; they just won’t get my data.
I could do all of this myself. I wrote those articles, I remember them. But Hypata remembers exact dates and connections better than I do, and it does the job in three minutes instead of an hour. Without it, I’d be sitting at DuckDuckGo doing the same thing more slowly.
I build the factual skeleton of the article myself. I decide that we’re recapping a year of Meta’s model development, from Llama 4 through LeCun’s departure to the acquisition and its Chinese chief, whose name I can’t spell. It’s like dictating an assignment to a senior colleague with an encyclopedic memory. I tell them where I want the article to go, and they bring objections and data.
Quite often the data breaks my thesis. Recently I wanted to use the saying “Prague’s dormitory” in campaign material for the municipal elections in Brandýs (a town just outside Prague). It sounds natural, and the saying is alive among Brandýs locals. But the commuting numbers showed a contradiction: almost as many people commute into the town for work as leave it for Prague. I almost wrote it without thinking, because everybody says it. This is the part of the work that AI does more rigorously than I do.
The second thing is technical inserts. Without Hypata I wouldn’t cover mobile networks at all. When a new feature arrives in a 3GPP release, you have to list the frequency blocks, modulation, and configuration exactly. None of the readers will be designing a phone, and they won’t remember the exact numbers. But if you get them wrong, the article loses its value even when its message is right. I used to download documents from the 3GPP.org FTP server and search through them. Today Hypata prepares this skeleton for me, and I put the meat on its bones, meaning the explanation that people come to my articles for.
You’ll object that AI hallucinates. Google “hallucinates” too (though technically it’s not the same). It hands you a document in which the information, for some reason, isn’t true, or it truncates a quote so that from the search preview you understand the exact opposite. Your senior experience sometimes saves you and sometimes doesn’t. The difference between the two tools is in speed, not in trustworthiness.
At the end I let the model go through the whole article. It checks for logical gaps, contradictions between individual facts and the overall message, and also for vulnerability to challenge, including legal challenge. After my article on the blackout, several ruffled parties spoke up. A more precise wording helped close that dispute.
Fine, but what does the reader get out of it?
This is an objection every author has to put to themselves, because everything above is a description of my own convenience. The reader doesn’t care whether I saved an hour.
The answer is that the reader buys a result and accountability for it. When you ask a junior colleague in the newsroom to verify the numbers, you don’t report that to the reader, because your name is still under the article and complaints come to you. With AI nothing has changed except the speed and the price of that junior. I’m the one who chose the topic, set the message, accepted or rejected the objections, and signed it.
The reader does lose something here, and it’s only fair to say so. When you leaf through a specification by hand, you occasionally stumble on something you weren’t looking for, and it becomes the point of the article. A targeted query to a model switches that chance off. The second loss is stylistic: I’m starting to adopt constructions that used to be written only by machines. Recently I wrote, on my own, where the company captures value, which is borrowed from the English value capture. That’s why I run a humanizer on the final text, a skill that looks for patterns people associate with AI. Typically the escalating argumentation that models drag into Czech from American journalism. Czechs tend to be touchier about this than publishers think, and the number of people who have an eye for it keeps growing. But something like this had happened to me before: in proper Czech one should write “sítě GSM” (networks GSM), but I traditionally write “GSM sítě” (GSM networks), the way it’s written in English texts.
What the law says
The rules under Article 50 of the AI Act apply from August 2, and the obligation to label text applies to content published for the purpose of informing the public on matters of public interest. But there is an exception that hardly anyone talks about: it doesn’t apply if the text has undergone human review or editorial control and someone bears editorial responsibility for the publication. So Europe answered the question from the opening of this article before I did. What decides it is responsibility for the result, not the machine’s share in producing it.
However, the interpretation adds two conditions, and my workflow violates them. A superficial grammar proofread doesn’t count as review. And if AI touches the text only after editorial approval, the exception falls away. I run the humanizer and the proofread as the very last step, and often AI also deploys the texts to the server. Either I reorder the steps, or the exception doesn’t belong to me. I haven’t decided yet. I have the footer on the site either way.
Why we’re returning to names
Two hundred dollars a month for a few dozen agents does work for which we had several dozen people at the Mobil Server newsroom twenty years ago. Without that, Vibecoding.cz wouldn’t work economically. I can’t afford to spend a whole day on a single article.
The change of format goes hand in hand with this. I notice myself that long, detailed texts are losing their audience, so I’m starting to lay out the most important points in two paragraphs at the beginning and only then argue them in detail. I often introduce that second part with a note that it’s optional, just for those who are really interested. It works, and it scares me a little. The reader feels they got a quality text because they can see they could read the details if they wanted to. Those other twenty paragraphs serve them as proof of my authority, not as reading material. They didn’t have that much time for the topic.
That brings us back to trust in names. I don’t know Mr. Tolasz personally, but I know his views on climate well enough that I accept his abbreviated claims without wading through the justification. This is exactly what will be more valuable in the age of AI. Elen Kovalčíková argues that the debate with people who have never come into contact with artificial intelligence will have value. I’ve also heard about demand for books published before 2022, because they don’t contain the patterns we now recognize as machine-made. I don’t have numbers for that, so for now I take it as an anecdote.
The change in itself is neither good nor bad, and asking the question that way is a mistake. After Gutenberg, books became cheap enough that it paid off to print in national languages for burghers without a Latin education, and writing adapted to that. Now it’s adapting again.
I dictated this text into Wispr Flow while cleaning at home. Otherwise I wouldn’t have written it. Then Hypata cleaned it up, because a stream of thoughts into headphones looks different from a stream of thoughts into a keyboard. The watermark is probably still in it. I’m the one who signed it. And Hypata put the referral into the Wispr Flow link, I see (and now I also see that it fixed my “refferal”).