CLAUDE’S INVISIBLE WATERMARK ON EVERYTHING AI WRITES
Claude, invisible watermark on everything AI writes. Wait a second, because we need some clarity here, because as often happens, the various influencers have spread wrong information. They don’t do it out of malice, but unfortunately AI is a complex world that takes skills going well beyond being nice or speaking well on video.
So what happened? Since the beginning of August, Claude leaves an invisible watermark inside every text it writes. You can’t see it when you read. There are no hidden characters and nothing is added to the end. It sits inside the way it picks words. As you know, a model writes one word at a time. The weather today was cold, and you could have “grey” or “overcast”. It barely matters. And normally the choice is settled by a random number. Now that randomness isn’t random anymore. A key decides it. Whoever knows how the key thinks, and it’s not public yet, knows how likely it is that Claude passed through there. Whoever doesn’t have it reads a perfectly normal text, and it slows nothing down. It costs no more and it adds not a single token. And they didn’t even invent the method now. It’s actually a version of SynthID-Text, published by Google DeepMind in Nature in 2024.
But why did they do it? Because Europe asks for it. As you know, I worked on the creation of that code of practice myself. And Anthropic applies it worldwide because it doesn’t yet have a clean way to limit it to Europe. And the other big ones will do the same.
But you see, there’s a very big but. The “watermark” only says Claude passed through there. And it doesn’t tell “Claude wrote it” from “Claude corrected it”. And that’s obviously a problem. If you write the text yourself and only ask Claude to fix it a little, or to translate into another language, maybe it’s the text that we actually wrote.
And when it arrives it’ll be a detection API, and nobody knows who will get access to it. What we’ll need is a procedure to challenge a wrong result. A student, a journalist, or an employee accused by a number nobody can explain is a problem. Because these things work only if verification is in the hands of both sides. The one accusing and the one being accused. What do you think?
#ArtificialDecisions #MCC #Watermark #Transparency #AIAct