Claude Opus 4.7 identified journalist Kelsey Piper from 125 unpublished words across multiple genres, signaling the practical end of prose anonymity for prolific writers.
Key Takeaways
Opus 4.7 matched Piper from a 125-word unpublished draft, a school progress report, movie reviews, a fantasy novel excerpt, and a 15-year-old college application essay.
The deanonymization is corpus-dependent: people with no substantial real-name writing online were not identified; the attack requires a large public reference corpus.
Post-hoc justifications from the model were often fabricated; the underlying stylistic fingerprinting works even when the explanations are nonsense.
Subculture proximity leaks identity: Claude named close friends of a target when it could not name the target directly, because shared writing communities share style tics.
Practical threshold: Glassdoor-style anonymous reviews may be attributable within one to two years as training data expands and models improve.
Hacker News Comment Review
Multiple commenters independently confirmed the effect on their own unpublished writing, including a nonfamous blogger and a fiction writer whose pieces were attributed to Greg Egan and China Mieville rather than themselves, suggesting Opus 4.7 maps to the nearest well-represented author when an exact match is absent.
There is skepticism about whether this is new: one commenter reported that raw GPT-4 pre-instruct was already completing text in a known physicist’s voice and signing his name years ago, implying instruct fine-tuning may have suppressed but not eliminated the capability.
One commenter noted that print-only regional authors were unidentifiable, reinforcing the corpus-size dependency and suggesting offline or low-digitization publication as a residual anonymity hedge.
Notable Comments
@dovin: Fed his own most-read blog post and got “Kelsey Piper” back, suggesting the model may over-index on certain writers as a catch-all attribution.
@chewxy: Unpublished fiction attributed to Greg Egan and China Mieville; accompanying blog posts attributed correctly to himself, showing fiction vs. nonfiction register affects match accuracy.
