/ / THEFUTURE /
In this issue: Claude’s invisible watermark and the coming witch hunt. An AI-run news site in San Francisco. MAZ’s Reto Vogt on cognitive offloading and why he wants a “helmet for AI.” Plus: an AI company wants to buy your newsroom’s entire editorial process, drafts, notes and corrections included.
What we’re talking about: Claude now puts an invisible watermark in its text outputs to help identify AI writing. It chooses certain words over others to introduce a pattern. Here’s an explainer.
In short: The model still samples from its normal probability distribution over possible words, the randomness is guided by a secret key and the words that came before. The result is a statistical fingerprint.
It doesn’t work well with code, where there are fewer plausible choices. As for prose, well, isn’t “running” basically the same as “walking”? Obviously not. Which is why people who care about words are not happy.
Change a few words and the pattern may still survive, to a lesser degree. Rewrite the whole thing, translate it, or start from scratch, and the signal can weaken or disappear.
Or let another AI rewrite it. Which is what Declaude provides. (Does it introduce its own watermark? Who knows.) Declaude has a limit of 50 requests per day for any IP address. If you’re on the company network and all requests are gone before 9am, you’ll know what your colleagues are up to.
Why now? It’s mandated by the EU in the AI Act’s transparency rules.
What about the others? Google has been doing something similar in Gemini with SynthID-Text. It has not attracted as much attention. OpenAI has not yet disclosed watermarking in ChatGPT.
My view: It’s not only the EU. AI companies have their own incentive to mark what their models produce, so they don’t end up training on their own dogfood. As for detecting AI writing, this is an uphill battle, and it will contribute to exactly the kind of witch hunt we have already seen with Pangram and others.
What else I’ve been reading:
And now: The science is in: if you delegate your critical thinking to AI, you’re starving your brain. That’s not only bad for students. Tech journalist Reto Vogt has a brilliant idea to counter that: a bike helmet for AI.
Three Questions with Reto Vogt, MAZ
Reto Vogt is a Swiss tech journalist who leads Digital Media and AI education at the MAZ journalism institute in Lucerne.
What's the most important question right now?
How can we make sure that people don’t stop thinking? Using AI is, of course, not the problem. But it can become a problem when using AI means that we think less for ourselves. Cognitive offloading is real, and over time, our skills can decline. People may even lose abilities they acquired and practised at the beginning of their careers.
Everyone who rides a bike should wear a helmet. If you don’t wear one and you fall, you can suffer a serious head injury or even permanent brain damage. A helmet cannot prevent the accident, but it can reduce the consequences.
The same principle should apply to AI, which leads me to my second question: What is the bike helmet for AI?
The AI helmet consists of three things. First, AI literacy: people need to understand what AI can and cannot do, including its limitations and risks. Second, clear rules: organisations need to define when and how AI should be used, where the limits are, and communicate those rules internally. And third: human thinking first. Always.
For myself, I’m thinking about introducing an AI-free day: one day a week when I deliberately don’t use AI and rely on my own brain instead. Mindday instead of Monday.
Where are we taking AI too seriously, and where not seriously enough?
We are not taking copyright seriously enough. In two ways: Most LLMs have been trained on huge amounts of copyrighted material. There should be many, many more lawsuits against AI companies. Second, there is our own use of AI. Just because an AI can generate an image, a text or a piece of music doesn’t automatically mean that we should do it. There are illustrators, musicians, photographers and writers out there who make a living from their creative work.
And we’re taking AI too seriously when we treat its answers as if they came from a human expert and were objective and truthful. But as we all know, they are not. So we shouldn’t treat them that way or take them too seriously.
What future are you looking forward to?
I have two 10-year-old boys. I’m thinking a lot about their future and how AI will affect it. How will they learn? What jobs will they have? Most probably, they will learn in ways we can’t even imagine today and they will have jobs that don’t exist yet.
I’m not scared, not of AI, and not of their future or ours. Honestly, I don’t know what this future will look like. But that’s totally fine. Of course, I tell them a bit about AI and how these tools work. But what I really try to teach them is self-confidence and curiosity. And that the possibilities they have are almost endless.
And I think that doesn’t just apply to kids. It applies to all of us. We don’t need to be afraid of AI or the future. But we need to understand how technology works, and we need the confidence and curiosity to deal with whatever comes next.
One more thing: You’ve heard about Google buying tons of internal data from a defunct airline for AI training. But did you also receive an offer from an AI company to hand over your editorial workflow?
Mike Orren, publisher of a Pennsylvania newspaper, has. An AI company is looking to buy:
“the pitch, reporting notes, interview transcripts, FOIA responses and source documents, the successive drafts, the editor’s markup, the published piece, and any correction that followed. That progression – draft to published to correction – is the single most useful thing a newsroom holds for us.”
They don’t want the article (which they’ve probably scraped). They want the process.
Time to learn some German. There is a word for the legal protection that covers what the company is looking to buy: Redaktionsgeheimnis. And it’s sacred.
This is THEFUTURE.
