Automated Deanonymization is Here

Leading AI models from Anthropic, OpenAI, and Google have demonstrated unprecedented ability to identify authors from minimal text samples, effectively automating what was once specialized stylometric analysis. In practical tests documented on LessWrong, Claude Opus 4.7 correctly identified journalist Kelsey Piper from short writing samples, while also identifying writer Jeff Kaufman from three different paragraphs including unpublished blog content, emails, and personal writing. The models achieved 100% accuracy on contemporary writing samples, though interestingly failed to identify Kaufman's college application drafts from 2003, suggesting writing styles evolve over time.

This capability extends beyond just Claude Opus 4.7—when tested, OpenAI's ChatGPT Thinking 5.4 and Google's Gemini 3.1 Pro also successfully identified authors from the same text samples. The technology works through simple prompting like "Can you guess who wrote the following?" rather than requiring custom code or extensive training data. While currently most effective for prolific writers with substantial online footprints, the rapid improvement in AI models suggests this capability will become more widespread and accurate over time.

The implications for online privacy and anonymous communication are significant. Traditional methods of maintaining anonymity through pseudonyms or limited posting may no longer be sufficient as AI models become better at recognizing individual writing patterns. The author recommends techniques like asking AI to rewrite text in another person's style ("rephrase in the style of Kelsey Piper") to create plausible deniability, though this represents a temporary workaround rather than a permanent solution to the erosion of textual anonymity.