The Washington Post published on Sunday an essay worth reading: modern models can identify an author by writing style, even on unreleased text. In the cited test, Claude Opus 4.7 nailed the author of a 1,000-word heist scene from a novel that has not even been published yet.
Why it matters
The classic New Yorker joke — "on the internet, nobody knows you are a dog" — just earned a giant asterisk. What used to require a forensic linguist billing by the hour now runs in seconds in a chat window. Every blog post you wrote, every commit message, every forum comment, every leaked DM became a biometric signature you cannot rotate.
- Frontier LLMs do stylistic echolocation with rising precision
- The Post's example used UNRELEASED text — i.e. it was not in the training data
- Career pseudonyms (writers, activists, whistleblowers) are at real risk
- "Defensive style" tools (paraphrasing to mask) are still immature and degrade text
- The regulatory discussion in Brazil today is essentially zero
What this means for creators
If you publish anonymously, semi-anonymously, or split your technical-work name from your personal-life name — this is a concern. Not theoretical. Anyone with access to a decent model and your public corpus can cross-link you with any new piece of writing on any platform. This applies to ghostwriters, devs maintaining a secret side project, content creators with alt accounts, and people commenting on politics who do not want to surface.
There is a practical path: run your own text through a local model before publishing and ask it to rewrite in someone else's style. Not bulletproof, but it raises the bar. The other good practice is not to cross-pollinate corpora — if the main account writes long and technical, the alt writes short and casual. Big models still get confused when registers are clearly distinct.
The real read
Online anonymity was always more fragile than it looked. AI did not invent the problem, it industrialized it. Anyone with a legitimate reason to write under a different name needs to start treating style as sensitive data — the same way we treat passwords. And anyone who has never thought about it should at least know the game changed.
Sources
- The Washington Post (April 26, 2026): Opinion — Artificial intelligence could kill anonymity online — https://www.washingtonpost.com/opinions/interactive/2026/04/26/artificial-intelligence-could-kill-anonymity-online/