Claude Opus 4.7 Really Does Know Kelsey

An experiment has been reported in which Anthropic's Claude Opus 4.7 accurately pinpointed a specific author, Kelsey Piper, from nothing more than unpublished short drafts. This result held true consistently across both incognito mode and API environments.

Opus 4.7 repeatedly identified Piper not just from texts on similar topics, but also from texts completely different in genre and time period, including movie reviews, fantasy fiction, and a college application essay from 15 years ago. This stands in contrast to ChatGPT and Gemini, which guessed other individuals.

The reasoning the model gave for its identifications was often logically inconsistent. This suggests that rather than reasoning like a human detective, the model is picking up on hard-to-detect stylistic tics.

This capability poses a real threat of loss of anonymity for people who have a lot of writing publicly available on the internet. While the model cannot currently identify everyone from a single paragraph, it has reached a level where it can narrow down the field to people around the author or members of a specific community.

In conclusion, as AI models continue to advance, the amount of data needed for de-anonymization via stylistic fingerprinting is likely to keep decreasing, and deliberately varying one's writing style may become necessary to preserve anonymity.