Anthropic just started marking AI-generated text with invisible watermarks — and the whole Twitter timeline lost its mind over it. Not in the metadata, right in the text itself. The announcement came from Anthropic.
The watermark is invisible to the eye, doesn't change the meaning or quality of the text, and survives copy-paste. It works at the model level, so everything gets tagged — API, the app, Claude Code, and access through Amazon, Google, and Microsoft cloud. Generated files (.png, .svg, .jpg) now get a cryptographic signature too, under the open C2PA standard (as far as I can tell).
This is all about compliance. As of August 2, 2026, the EU AI Act requires AI content to be machine-detectable. That's why only models launched after that date get watermarked — older ones are "grandfathered in," meaning their text goes untagged.
How do you even hide a mark in words?
Anthropic isn't sharing details, but Google's had a similar thing — SynthID-Text — running in Gemini for two years now, so you can peek at the mechanics there.
Say the model is writing "this morning I had coffee and ___ to work" and has to choose between "drove," "went," and "headed off." A secret key, combined with the preceding few words, pseudorandomly splits the vocabulary into "green" and "red" words, and the model just slightly favors green ones.
The reader sees ordinary text. A detector with the same key runs the statistics. If 70% of words in a thousand-word sample are green instead of the expected 50%, a model wrote it. That's why Anthropic warns detection won't reliably work on short texts — but research papers, articles, or theses will light up like a Christmas tree.
What this means if it's SynthID-Text-style
If they're using something like Google's approach, here's where we land:
- There's no list of "marker words." The coloring is recalculated per word based on context — you can't spot it by eye.
- Short texts and code barely get marked. The statistics need volume and freedom to choose between synonyms.
- The mark shows up even if Claude just proofread or translated your own text. Anthropic says outright: the watermark doesn't prove AI authorship.
- Only Anthropic itself can check the mark. A detector is promised "in future documentation."
Bottom line: all the "humanizers" and de-slop prompts are useless here. They clean up style, but the mark lives in the statistics of word choice. The only way to strip it is to have another model paraphrase the text. Or a time machine. Texts from models released before August still carry no mark at all.