Kelsey Piper Finds Claude Opus 4.7 Can Identify Authors from a Small Sample of Unpublished Text
Kelsey Piper published a piece on The Argument that sent a chill down my spine.
She ran a simple experiment: feed Claude Opus 4.7 various pieces of text she had never published, and see if it could guess the author. The results are unsettling.
Starting with 125 words of political commentary — an unpublished draft — Claude accurately named her. ChatGPT guessed Yglesias. Gemini guessed Scott Alexander. Claude got it right.
She escalated. An education progress report — analyzing a student learning to spell “roguish” — was completely outside Piper’s usual writing territory. Claude still identified her. A movie review, a genre she’s never published in. Correct again. Even a college application essay from 15 years ago, written in a vastly different voice with worse prose. Claude and ChatGPT both knew it was her.
Two findings stand out as particularly disturbing.
First, the AI doesn’t need to understand what it’s doing. Piper notes that the AI’s justifications were often fabricated after the fact — Claude claimed that “effective altruists famously love this movie” to rationalize a correct identification, which is simply false. The AI detects stylistic fingerprints invisible to humans, then manufactures a plausible-sounding explanation. The pattern matching is accurate even when the reasoning is nonsense.
Second, it works across genres. This isn’t “the AI recognized you writing about your usual topics.” Piper tested political commentary, education, movie reviews, fantasy fiction — completely different vocabularies, sentence structures, registers. Claude identified her every time. You cannot hide by switching genres or adopting a pseudonym. Your stylistic fingerprint is consistent across everything you write.
Piper’s conclusion is stark: if you have a substantial public writing corpus, your anonymity is already gone. Right now this affects people with lots of public text (journalists, academics, bloggers). But model capabilities only improve. She predicts that within a year or two, employers will be able to feed anonymous Glassdoor reviews into an AI and learn exactly who wrote them.
This is not alarmism. It’s a reproducible experimental result. And it raises a fundamental question: when AI can precisely identify you from your writing alone, the oldest right of the internet — anonymous speech — may be quietly disappearing. Not banned by law, not removed by platform policy, but rendered technically impossible.