AIが書いた文章、どうやって見分ける？

Claude's August 2026 Watermark Announcement

In August 2026, Anthropic announced that it had implemented a watermarking feature for text generated by its AI service Claude, sparking widespread discussion across the AI community. Unlike a visible stamp on an image, the watermark is undetectable under normal reading conditions, but a specific method can identify the text as Claude-generated.

Why Watermarking Text Is So Hard

Images and PDFs exist as files, which means they have a dedicated area for "metadata" — the date a photo was taken, the camera model used, and so on. But when you copy and paste text that Claude has produced and post it to social media, that text has never existed as a file. Without a file container, there is nowhere to attach metadata, and no obvious place to hide a watermark.

Another candidate approach is the "zero-width space" — invisible special characters inserted within the text that survive copy-and-paste and can later be detected. The problem is that anyone with the right knowledge can remove them trivially. A technically minded person can spot the extra characters and delete them instantly, and a dedicated removal tool would be written almost immediately. For that reason, this approach was not adopted.

How Claude's Watermark Actually Works

The method Claude uses is to subtly manipulate the probability with which words are chosen during generation. AI produces text by probabilistically selecting the next word based on the preceding tokens (roughly, words). Anthropic adjusts those probabilities by a tiny amount.

Picture a sentence where either "gray" or "mouse-gray" would work equally well. The system is configured so that "mouse-gray" becomes ever so slightly more likely to be chosen. It does not always pick "mouse-gray" — text quality is unchanged. In a context where Hercule Poirot's tagline virtually demands "little grey cells," mouse-gray will not appear. The bias only activates in genuinely interchangeable moments.

What makes the technique especially sophisticated is that the bias is not a fixed word-level preference but a context-dependent one. The value derived from mathematically processing the preceding four tokens determines how the next word's probabilities shift. The phrasing "today from morning the clouds" might nudge the choice toward gray, while the slightly different phrasing "today from morning somehow the color of the clouds" nudges it toward mouse-gray. Because no human reader can perceive any pattern in the text, simply replacing every instance of "mouse-gray" with "gray" cannot strip the watermark out.

The algorithm that produces this bias is publicly documented, but the secret key used inside the process is held exclusively by Anthropic — making outside imitation or reproduction impossible. The background to this implementation is the EU AI Act, which mandates the identification of AI-generated text.

How It Differs from Popular AI Detection Tools

Search online and you will find many tools that display something like "this text has an X% probability of being AI-generated." These work on an entirely different principle. They learn patterns of vocabulary, writing style, and logical structure that AI commonly produces — polished word choices, a habit of presenting both sides of an argument, a tendency toward safe and palatable answers — and use those patterns to infer "AI-likeness."

Their accuracy has clear limits. In the episode, Mizuno-san (水野さん) ran an essay she had written herself through an AI checker and received the verdict "15% AI" — a straightforward false positive. Tools like these operate at the level of a police officer acting on a vague hunch; wrongful arrests (false detections) happen.

Claude's watermark, by contrast, can identify a sufficiently long piece of text as Claude-generated with 99.7% confidence — the equivalent of DNA evidence. However, Anthropic has not yet made either the secret key or the detection tool available externally, so in practice there is currently no way for a university professor checking a student's report, or an editor trying to spot an AI manuscript, to actually use this capability.

The Paradoxical Conclusion: "How to Submit an AI Report Without Getting Caught"

Because the watermark is context-dependent, rewriting scattered portions of the text causes detection accuracy to drop sharply. Since the mechanism looks at the preceding four tokens when selecting each word, replacing roughly one word in four would erase the bias at most positions and make detection very difficult.

But if you are going to put that much effort into reworking AI-generated text, you would have been better off writing it yourself from the start. The episode lands on a paradoxical conclusion: if you absolutely do not want to be caught, writing it yourself is the recommended approach.