Claude Opus 4.7 identified Kelsey Piper as the author of unpublished drafts across genres from as few as 125 words, in incognito, via API.
Key Takeaways
Opus 4.7 correctly named Piper from a 125-word political draft, an education progress report, a movie review, fantasy fiction, and a 15-year-old college application essay.
The capability is Opus 4.7-specific for now: ChatGPT and Gemini guessed wrong on most tests; Opus 4.6 also failed on some samples.
Post-hoc explanations from the model were fabricated rationalizations; the underlying stylometric signal is real even when the stated reasoning is nonsense.
Anyone with a large public real-name corpus is currently deanonymizable; those with no significant public writing are not yet at risk.
Piper predicts deanonymization thresholds will drop over time, eventually covering Glassdoor reviews, Discord posts, and other low-volume anonymous text.
Hacker News Comment Review
Commenters split on whether this is genuinely new: one noted that raw GPT-4 pre-instruct already completed text in a known physicist’s voice and signed his name, suggesting stylometry has been latent in large pretraining corpora for years.
There is skepticism about the mechanism: the stylometric task here (identify a specific blogger) is harder than LLM-detection, yet no prior model matched it reliably, leaving commenters uncertain what changed architecturally in Opus 4.7.
Independent replication was fast and positive: at least two commenters fed their own unpublished drafts in incognito and got correct identifications, and one noted combining multiple passages with contextual signals would push accuracy far higher.
Notable Comments
@Extropy_: Suggests feeding the Bitcoin whitepaper to see who Opus 4.7 names as author.
@sodacanner: Points out chain-of-thought logs would be more revealing than post-hoc explanations; the article only tested asking after the fact.
