AI Identifies Authors From 150 Words

Updated: 2026.04.27 1M ago 2 sources
A current-generation LLM (Anthropic’s Claude Opus 4.7) can attribute short, unpublished text excerpts to a real individual reliably from roughly 125–150 words, even across registers and drafts. The capability works without account memory and in Incognito or API settings, meaning stylistic fingerprints alone can suffice. — If widespread, this capability undermines online anonymity and will reshape debates about free expression, whistleblowing, platform policy, and legal protections for anonymous speech.

Sources

Will AI end anonymity?
Tyler Cowen 2026.04.27 95% relevant
Cowen provides on‑the‑record, quantified examples: Claude (Opus 4.7) identified him from as few as 124 words and identified other unpublished chapters in ~1,100–1,400 words, directly illustrating the existing claim that modern LLMs can deanonymize authorship from very small text samples.
I can never talk to an AI anonymously again
Kelsey Piper 2026.04.21 100% relevant
Anthropic’s Claude Opus 4.7 identified Kelsey Piper from a 125-word unpublished column excerpt and from other unpublished school-report drafts, tested in Incognito and via API.
← Back to all ideas