Claude 4.7 Identifies Journalist from 125 Words

When Kelsey Piper fed 125 words of an unpublished political column into Claude 4.7, the model guessed her name.

• She reduced the chance of account-based identification by using a logged-out state and API calls, and got the same result on a friend's laptop.
• Even with completely different genres — a school progress report and a review of a 1942 war comedy film — Claude kept pointing to her.
• In the same test, ChatGPT and Gemini both gave wrong answers.

This result suggests that an author's voice may not just be a matter of stylistic preference, but a multi-axis fingerprint that models can read.

It also points to the possibility that Anthropic's models have a stronger ability to read prose patterns than other models, which invites the interpretation that this ability may be linked to stylistic rigidity at the generation stage as well.