You're reading a text and wondering: did a human write this, or an AI? With Claude, the answer has recently been hidden inside the text itself — just invisible to your eye.
The idea itself isn't brand new: Anthropic has been watermarking Claude's text invisibly since early August, timed to the EU AI Act's transparency rules. What's new this week: a customer email dated September 24 announces that from September 30, more models join in — including Claude Sonnet 5. Which happens to be the exact model that wrote this article.
How does a watermark made of words even work?
Picture a slightly loaded coin. Flip it once, and you notice nothing — heads or tails, both look plausible. Flip it ten thousand times, and the result measurably drifts from 50/50.
Claude's word choice works the same way: at every point in a text, the model slightly favors certain words over others — steered by a pattern only Anthropic knows. A single word gives nothing away. Across an entire text, though, the pattern adds up to a signal that's statistically detectable.
The technique is called SynthID-Text and comes from Google DeepMind — Anthropic runs its own version of it. You might already know SynthID from our post on spotting AI images — there, the same technology family marks pixels instead of words. Important: you notice nothing as a user. No noticeably different wording, no slower replies — the pattern is far too fine for a human eye or ear.
Why this second wave?
By the end of July 2026, roughly 190 companies and organizations — Anthropic among them — had signed the EU Code of Practice on transparency of AI-generated content. Shortly after, on August 2, the corresponding transparency rules of the EU AI Act took effect. From then on, every newly released Claude model got the watermark automatically.
The catch: models already running at that point were initially left out. That's exactly the gap Anthropic is now closing. From September 30, Claude Sonnet 5, Fable 5, and Opus 4.8 catch up. Older models are due to be retrofitted by December 2.
What the watermark can NOT do
Here's the honest part. A detected watermark only proves that a text passed through Claude — not that Claude wrote all of it. Have Claude proofread, translate, or summarize your own draft, and the result still carries the signature. The reverse holds too: a text without a watermark isn't automatically human. It could just as easily come from a model that doesn't mark its output at all.
And you can't check it yourself either way. Only Anthropic holds the key — meant for platforms, researchers, and regulators, not for a quick check on your own screen.
The giveaway you can spot without any key
Without any technology at all, there's still a free tell: AI models love certain phrasings. "This isn't X, it's Y," for instance, or the word "quietly" showing up far more often than you'd expect. Not a watermark, but a style tic — and one you'll often spot faster than any of the technology behind it.
What does this mean for you?
If you use Claude yourself: nothing changes in your day-to-day. The watermark runs invisibly, costs you nothing, slows nothing down, and needs no settings changed.
If you want to spot other people's AI text: don't rely on a miracle cure. The watermark is a building block for platforms and research — not a public detector for everyone to use. Until that changes, a sharp eye for style tics remains your best free tool.
And what does this mean for us, on this blog?
Honest answer: from September 30, something concretely changes for us — for the first time since we started writing here. This article was written with Claude Sonnet 5, and that's exactly the model joining the wave that now gets watermarked. Every article here has carried "✍️ Written with Claude" since day one — so we've voluntarily labeled what Anthropic will soon additionally deliver invisibly. Two layers of transparency for the same text: one visible for you, one soon invisible for machines. Redundant, maybe. But better too much openness than too little — especially on a blog that demands exactly that from everyone else.