Transcription

• Speaker #0

  Next time, the ones who hold the keys. That was the promise we left you with last week. This week, the key turns out to be real. In a few weeks, across the European Union, text written by ChatGPT will carry an invisible mark, woven into the very choice of its words. You won't see it. And you won't be able to read it. To read it, you need a key. And for now, that key is only available to a handful of approved researchers. That same week, OpenAI's newest model reached every account, free ones included. And Claude, Anthropics Assistant, moved into your Google Documents. So when a machine writes alongside you, who can say who wrote it? Welcome to the week of October 5th, 2026, Synth AI Cast, the AI that listens for you. Episode 28, the 6th of Season 2. This week, a mark nobody can see, and two keys that are not in the same hands. If this sounds familiar, it should. Back in August, in Episode 20, we looked at Anthropic's mark. Today, OpenAI adds its own, and we compare the two approaches. Let's start with the mark. On Monday, OpenAI unveiled something called TextGrain, a watermark for text produced by ChatGPT, and by Codex, its coding tool. The idea is simple to explain. Whenever the model hesitates between several possible words, a secret key gently nudges its choice. Nothing is added to the text, no hidden characters, but over a few hundred words. All those small choices add up to a fingerprint. The mark will be switched on by default, on every plan, but only inside the European Union. Everywhere else, nothing changes. Developers around the world can turn it on if they want to, and that's it. So why Europe? Because the EU's Artificial Intelligence Law, the AI Act, has required it since August 2nd. Its Article 50 says generated text must be recognizable by a machine. Tools already on the market have until December 2nd to comply. In practice, for you. A school essay. A cover letter. A message to a client. Written with ChatGPT. from inside the EU will carry that fingerprint. And here is something unusual. OpenAI publishes the limits of its own mark on passages of 400 tokens. Its detector recognizes the mark about 92 times out of 100. Swap one word in 10 for a synonym, and that drops to 66. Swap one word in 4. and it finds the mark less than one time in five. Short texts are harder to catch, and so are texts where there are few words to choose from, like mathematics. The company spells it out. The absence of a mark does not prove that a human wrote the text, and its presence does not measure how much of the work was human. On Wednesday, a French podcast, Le Centre de Gravité. Asked the question in its title, can the watermark prove who the author is? According to OpenAI itself, no. Two years ago, OpenAI already said it had a method like this. And it chose not to roll it out, partly for fear of penalizing people who write in a language that isn't their own. This week, the law made the call. For the first time. A European text is changing a product that millions of people use every single day. And the detector? For now, OpenAI is keeping it for researchers and expert organizations. Approved case by case, on application. According to OpenAI, a detection reveals neither your name nor your conversation. And the method is meant to be released as open source. with no date given. Here's a detail worth noticing. OpenAI's checker for images and audio is open to any organization. For text, It isn't. The key that marks belongs to the company. The key that reads is lent out. On application. Second signal. The comparison we promised. With Anthropic, the same principle. A fingerprint in the choice of words. Built on a method published by Google DeepMind. And like OpenAI, under the European Code of Practice. So what actually changes for you? First, the map. Anthropic marks text from its new models everywhere in the world. OpenAI, only in the European Union. If you're in Paris or Berlin, both assistants will soon mark your text. In London or New York, only one of them will. Next, the numbers. OpenAI publishes its own. With a technical report of about 20 pages, Anthropic's help page publishes no error rate. But on the essentials, the two are remarkably alike. At Anthropic 2, the detector is restricted. Regulators, law enforcement, the media, researchers, on request. Either way, you are not the one holding the key. Anthropic promises to widen that access over time. And on Tuesday, Claude moved into your files. An add-on for Google Docs, Sheets, and Slides. In public beta for every paid subscriber. It reads the open document and edits it right there. A grammar fix, a table, one more slide. Anthropic says so itself. Its mark signals text written. or simply processed by Claude. The more it writes in your sentence, the more room the mark has. So your report, corrected by AI inside your own file, whose is it? And since we're talking about transparency, the script I'm reading to you was written with Claude. It most likely carries that mark, and we can't read it either. Third signal, a shorter one. On Wednesday, OpenAI launched GPT-6 in ChatGPT for subscribers, and by Thursday, for free accounts too, a version called Sol for subscribers, and a lighter one, Luna, for everyone else, with illustrated answers, charts, buttons, small calculators. Just one day apart, the newest model now writes for everyone, and in Europe, within a few weeks. What it writes will carry the mark. More and more text will carry a mark. And very few people will be able to read it. Let me take you to Italy, at the end of the 13th century, around the year 1282. Paper makers start sewing a thin metal wire into their molds. The sheet dries and keeps a design that is thinner and lighter. It's the ancestor of what English calls a watermark. By the 15th century, almost every sheet of paper carries one. Historians still use them today to date old documents. To see it, you simply hold the sheet up to the light. Anyone can do it. For more than seven centuries, a watermark was read against the light. This one can only be read with a key. Back in August, we talked about the Goldsmith's Hallmark. A mark is only worth something if someone, open to all, can read it. This week, that someone exists, but you have to apply. Last week, we asked who holds the keys. The answer is that there are two. The one that marks, kept by the company, and the one that reads, lent out to a few. The mark was imposed by the AI Act. The keys are kept by the companies. One word in ten. That's all you need to change for the mark to fade sharply. That number doesn't say the mark is useless. It says it proves far less than people will think. What to watch now? The exact date the mark switches on. Where you live. The list of who gets access to the detector, and the day a school, an employer, or a platform, relies on that mark to settle a dispute. Three things to do this week. Five minutes. Check whether the assistant you use already marks your text. The answer is on its help page. 30 minutes. If you teach or manage people, re-read your rules on AI at work or at school, and ask yourself whether they treat a detected mark as proof. One hour. Read Article 50 of the European AI Act. It's already changing your tools. This week, the mark arrived before anyone knew how to read it. If a machine can tell that AI has touched a text, who will have the right to ask you about it? Next time, the ones who ask the question.