Opus 4.7 knows the real Kelsey
Advances in large language models are revealing an unexpected capability: accurately attributing anonymous or unpublished text to specific online writers based solely on style and content. Commenters report Opus 4.7 correctly identifying authors of drafts, blog posts, and chat messages, raising fears that stylometric deanonymization at scale could effectively end practical online anonymity. The thread explores how this might affect vulnerable groups, future privacy norms, and whether defensive tools — such as AI-based style obfuscation — can meaningfully mitigate the risk.
Perceived new capability: automated stylometry
- Many commenters report that the model can correctly guess the author of short text samples, including:
- Unpublished blog drafts and book excerpts.
- Private or semi-private community posts.
- Posts written after the model’s stated training cutoff.
- Others see clear limits: mislabeling ordinary posts as coming from a few prolific writers, or only narrowing to a “type” of tech/rationalist blogger.
Memory vs training vs genuine inference
- Multiple people stress that memory and account linkage were disabled or controlled (incognito, API, different users), yet the model still identified authors.
- Some speculate earlier testing texts may have entered later training.
- Several note that explanations for “how” the model recognized an author felt post‑hoc and implausible; the model likely can’t introspect its real mechanism.
Privacy and deanonymization concerns
- Strong theme: this looks like the beginning of routine deanonymization from writing style, especially for anyone with a sizable public corpus.
- Commenters fear:
- Linking pseudonymous posts or private emails to real identities.
- Outing vulnerable groups or political minorities at scale.
- Future models using personal AI chat logs to answer questions about individuals.
- Some argue effective online anonymity may never have truly existed, given infrastructure-level tracking.
Defenses and trade‑offs
- Proposed defenses:
- Run all writing through a local or separate LLM to “de-style” it.
- Intentionally write in a non‑native language or distorted style.
- Use stylometric encoders/decoders with trusted contacts.
- Many find these options distasteful or harmful to authentic human voice and discourse.
Broader social and ethical implications
- Some imagine a near‑zero‑privacy world: potentially safer (less hidden crime) but also more oppressive and dull.
- Several tie anonymity to protection for unpopular or stigmatized groups; others push back on how such examples are framed but not on the core privacy risk.
- There is both awe at the technical feat and alarm at its implications; uncertainty remains on how widespread and reliable this ability really is beyond heavily published authors.