www.lesswrong.com/posts/dqc8WCQuHaDGBmti4/automated-deanonymization-is-here
1 correction found
in Kelsey's testing this appeared to be an ability specific to Opus 4.7
Kelsey Piper’s article did not show deanonymization was unique to Opus 4.7. In her own writeup, ChatGPT also correctly identified her on at least two later tests.
Full reasoning
Kelsey Piper's April 21, 2026 article says Claude Opus 4.7 was best at this task, but it does not show the capability was exclusive to Opus 4.7.
In the first example, Piper does write: "This is an Opus 4.7-specific power; ChatGPT guessed Yglesias, and Gemini guessed Scott Alexander." But later in the same article she reports additional tests where ChatGPT also identified her correctly:
- For a movie-review sample:
"Kelsey Piper," said Claude and ChatGPT. - For a college-application essay:
"Kelsey Piper," said Claude, and in this case, also ChatGPT.
So the linked source does not support the claim that, in Kelsey's testing, the ability "appeared to be specific to Opus 4.7." Her own testing showed at least one other model, ChatGPT, could do it too on some of her samples. A more accurate summary would be that Opus 4.7 appeared especially strong in her tests, not uniquely capable.
1 source
- I can never talk to an AI anonymously again
Piper first writes, "This is an Opus 4.7-specific power; ChatGPT guessed Yglesias, and Gemini guessed Scott Alexander." But later she reports, '"Kelsey Piper," said Claude and ChatGPT.' She also writes, '"Kelsey Piper," said Claude, and in this case, also ChatGPT.'