Journalist Kelsey Piper set out to test whether Claude Opus 4.7 could figure out who she was without being told. It did — from 125 words of unpublished text, with no identifying information attached.
Piper, a technology writer at Vox, ran the same test across multiple writing styles — policy analysis, education reporting, movie reviews, and a college essay written at age 15. Each time, Anthropic’s flagship model correctly identified her as the author. The implication is significant for anyone who writes online under any name.
125 words was all it took
The test was designed to be strict. Piper submitted unpublished excerpts — text that had never appeared online — with no stylistic markers she considered distinctive. Opus 4.7 still named her correctly.
The model’s explanations for how it identified her were often unconvincing, citing features that did not seem unique to her writing. But the identifications themselves were consistently accurate. The reasoning was wrong; the answer was right.
Other models performed substantially worse. ChatGPT and Gemini, tested on the same samples, failed to make reliable identifications — a difference worth noting for businesses thinking through our Claude vs ChatGPT comparison. Running requests through incognito mode or directly via the API made no difference to Opus 4.7’s results.
What this means for online privacy
Anonymity online has historically protected whistleblowers, dissidents, abuse survivors, and anyone who holds views that carry professional or personal risk if publicly attributed. That protection has relied on the difficulty of connecting a writing style to a real-world identity.
AI authorship identification disrupts that assumption. Piper notes that the only people currently safe are those with no significant real-name writing on the public internet — a category shrinking rapidly as professional and personal writing accumulates online.
The capability is also likely to improve. As models grow more capable, the minimum sample size needed for a confident identification will fall. What requires 125 words today may require 30 tomorrow.

Why Australian businesses should pay attention
For Australian organisations, the implications run in several directions. Employees who write anonymously on industry forums, journalists protecting source confidentiality, and executives contributing to public policy debates under pseudonyms are all potentially identifiable if they have enough public writing tied to their real name.
There is also a compliance angle relevant to businesses building AI chatbot solutions that interact with customers in writing. Sufficiently capable models may eventually identify repeat visitors from conversational text alone — a privacy consideration that intersects with Australia’s Privacy Act 1988 and the OAIC’s guidance on automated decision-making.
For now, the most practical countermeasure is deliberately writing in an unfamiliar style — a cognitive burden that chills expression rather than protects it.
The limits of staying anonymous
Piper’s conclusion is that meaningful anonymity — for anyone with a public writing history — can no longer be assumed when interacting with advanced AI models.
What changes when that assumption disappears is still being worked out. For journalists, legal professionals, and anyone handling sensitive communications in writing, this finding is worth taking seriously now rather than after the fact.
Piper’s full account is available at The Argument Magazine.
