Is AI Disclosure Day here?

TL;DR: Anthropic will embed a new AI watermark in all text generated by Claude—and it may not be that easy to remove. It’s doing this to comply with an EU law that just went into effect, but it also lands as public fatigue around AI-made content reaches an all-time high.

What happened: This is good news for haters of AI writing, as the imperceptible watermark won’t get stripped out by copying and pasting.

Anthropic dropped the news yesterday, and it’s a glimpse of the impact the EU’s new AI transparency law might have. The law, which went into effect on August 2, requires providers of AI tools to mark their outputs in a “machine-readable format.” Anthropic is rolling out the watermark worldwide for all new models, and will eventually add it to existing ones.

The announcement comes as backlash against AI slop reaches critical mass, with platforms rolling out new features to label AI content seemingly every day. We don’t know exactly what Anthropic’s watermark will look like, but the system could be similar to Google’s SynthID, which works on text, images, videos, and audio. Another unanswered question: how detection tools will be made available to the public. Anthropic says it’s working on letting users and “other third parties” detect its watermark, as part of a voluntary Code of Practice the company recently signed.

How it (probably) works: AI text watermarks get embedded in the actual sequence of words—they’re not some invisible character you can remove by, say, manually retyping the text.

At a basic level, AI text generation works by predicting which word is statistically most likely to follow the previous one. But don’t worry, em-dash stans—your writing probably won’t get flagged just for sounding vaguely AI. These watermarks deploy a secret key that tweaks the probability of words used (leaving a signature that it was generated by a certain lab’s models)—and text that doesn’t match that pattern won’t carry the mark.

The big but: AI text watermarks aren’t tamper-proof. One study from Queen’s University found that paraphrasing with another AI model or using back translation—translating text into another language, then back to the original—could reduce detection of watermarked text. Or you could just dilute the AI text by dropping it into a larger body of human-written words so it’s harder to find.

Bottom line: Most major AI labs signed the same Code of Practice that Anthropic did—notable absences include Amazon and xAI—and everyone has to follow the new EU transparency rules. It could all usher in a drastic change for how AI text gets disclosed. And perhaps even how many AI slop books we see on digital shelves. —WK