Latest / Minds, Bodies, and Terawatts / Every Sentence Claude Writes Now Carries an Invisible Mark - and It Can't Tell Anyone Who Wrote It

Transcript

Every Sentence Claude Writes Now Carries an Invisible Mark - and It Can't Tell Anyone Who Wrote It · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, the show where the machine signs its own work in ink nobody else can read. Anthropic published a technical explainer on August fourteenth describing how it will stamp an invisible watermark into every piece of text Claude produces.
• 0:16
  [Guest] Worldwide, at launch. Anthropic says it can't yet scope the mark by region.
• 0:21
  [Host] So a European rule lands on my chatbot in Ohio because switching it off for Americans was too much trouble?
• 0:29
  [Guest] Roughly. The trigger is Article Fifty of the European AI Act, whose transparency rules took effect August second.
• 0:36
  [Host] How do you hide a watermark in a sentence? There's no pixel to nudge.
• 0:41
  [Guest] Word choice. Every time the model picks between synonyms, it tilts that coin flip toward a pattern set by a secret key. Invisible to a reader, readable to anyone holding the key.
• 0:53
  [Host] Good. So a teacher or a boss can finally prove somebody used AI.
• 0:58
  [Guest] With what? They don't have the key. Anthropic does. The only party who can check whether Claude touched your document is the company that sold you Claude, and it says a detection tool is coming.
• 1:12
  [Host] So the evidence is real and only the vendor can read it.
• 1:16
  [Guest] And even in Anthropic's hands it's blunt. Its explainer says a hit means Claude was likely involved at some point, not that Claude wrote the thing.
• 1:26
  [Host] Where does it break down?
• 1:28
  [Guest] Exactly where people get accused. Detection is weak on short samples, thinner on factual passages, and close to nothing in code, where there are few free word choices.
• 1:39
  [Host] Is that measured, or is that Anthropic hedging?
• 1:42
  [Guest] Measured. Kirchenbauer and colleagues at Maryland published "On the Reliability of Watermarks for Large Language Models" at ICLR twenty twenty-four. The mark does survive human paraphrase, but you need eight hundred tokens, roughly six hundred words, before the call is reliable.
• 1:58
  [Host] A cover letter is two hundred.
• 2:00
  [Guest] Which is the entire genre people write with a chatbot. And Sadasivan and colleagues, also at Maryland, in "Can AI-Generated Text be Reliably Detected?", showed a cheap paraphrasing pass drives detector accuracy toward a coin flip.
• 2:14
  [Host] So the incentives invert.
• 2:15
  [Guest] Light edits leave the pattern intact, a full rewrite strips it. The freelancer who had Claude tidy a paragraph stays tagged. The one who ran the file through a paraphraser comes out clean.
• 2:28
  [Host] Is anyone pushing back?
• 2:29
  [Guest] John Gruber, on Daring Fireball, called it a perversion of writing. Anything but his own needs shaping his words, he wrote, is patently offensive.
• 2:39
  [Host] And somebody's already built the eraser.
• 2:41
  [Guest] A developer named James Padolsey. His tool is called Declaude. Paste text in, get plain prose back, and Gruber says the mark goes with it. His verdict: weak watermarks appeasing a weak law.
• 2:54
  [Host] So the honest get flagged and the cheaters strip it. Why ship it?
• 2:59
  [Guest] Compliance, stated plainly. Google's Gemini has watermarked text since twenty twenty-four; OpenAI hasn't published a text plan at all.
• 3:06
  [Host] Except we've run this experiment. Turnitin's own site says its AI writing detection may not always be accurate, so it shouldn't be the sole basis for action against a student.
• 3:17
  [Guest] And schools used it anyway.
• 3:19
  [Host] So what's genuinely new here?
• 3:21
  [Guest] The first checker for machine writing at real scale, and it checks the wrong column. Not whether the work is any good. Only where the words came from.
• 3:32
  [Host] Patrick Deglon has a Law of Gravity in Unscarcity called Truth Must Be Seen. No decision in the dark, because you can't appeal what you can't inspect.
• 3:42
  [Guest] Anthropic shipped a transparency measure only Anthropic can see through. For more on proof, provenance, and who gets to check your work, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.