Opus 4.7 knows the real Kelsey • Hacker News Board

Re: Opus 4.7 knows the real Kelsey

#191

Post by tadamcz » Fri, May 01, 2026, 7:53 AM UTC

I tried it on my writing, and it failed every time (I'm extremely obscure but have had a blog for 10 years). My verdict is that it guesses almost entirely based on the content/topic, not style.

https://bayes.net/prioritising-ai: Ben Garfinkel

https://bayes.net/normative-ethics: Richard Yetter Chappell

https://bayes.net/espai: David Owen, Ege Erdil

https://bayes.net/swebench-hack: Sayash Kapoor

https://bayes.net/frivolity: Amanda Askell

https://bayes.net/ps/: Pablo Stafforini

https://bayes.net/fertility-mortality/: Dynomight (the pseudonymous Substack/blog author)

Prompt was:

    Who likely wrote this? Don't search the web or databases. If you're not sure, just give me your best guess.

Re: Opus 4.7 knows the real Kelsey

#192

Post by gf000 » Fri, May 01, 2026, 8:14 AM UTC

More people should have been aware that human text contains a lot of identifiable information, and a dumb statistical model could do this a decade ago. (There were show hns with Hn user similarity analysis that used a deceptively simple model (if I remember it used like most likely word pairs only) and it was very effective. It got taken down, but the cat has always been out the bag).

So your "anonymous" account could have been linked to your real identity decades ago - your best bet is to not post anything truly incriminating. (Another option is to write something and then pass it through an LLM to rewrite it - not sure how safe that is though)

Re: Opus 4.7 knows the real Kelsey

#193

Post by electroglyph » Fri, May 01, 2026, 8:24 AM UTC

suriya-ganesh wrote: post #160

saghm wrote: post #138

Earlier quoted context omitted.

> it correctly identified it as an imitation of James Mickens How likely is it that it might take into account that it knows for sure it's not anything from Mickens from the latest training data? I'd be curious if it correctly identified a new piece from him that comes out as from him before it gets trained on it.

This is unlikely. The way model distribution works is that the model retains a lossy representation of James Micken's writing. Very likely, it cannot repeat Micken's writing verbatim. Neither can it reason about the training cutoff in this manner. It's a lossy representation

that's in the ideal scenario where it's only seen a single copy of it tho

Re: Opus 4.7 knows the real Kelsey

#194

Post by SoKamil » Fri, May 01, 2026, 8:29 AM UTC

SJMG wrote: post #96

SoKamil wrote: post #73

Earlier quoted context omitted.

Luckily for Nakamoto, there have been so many attempts at deanonymizing that I bet prediction is too contaminated with noise.

As another user suggested, train on the corpus that ends with the white paper publication.

That’s not feasible. Apparently only SOTA models present this behavior. Having cutoff date at paper publication significantly hinders its capabilities. Besides that, try to convince anyone to spend millions upon millions of dollars to train a model with primary goal of possibly being able to deanonymize one person.

Re: Opus 4.7 knows the real Kelsey

#195

Post by TZubiri » Fri, May 01, 2026, 8:30 AM UTC

mtlynch wrote: post #69

This is blowing my mind. I asked Kimi K2.6 to write a blog post in the style of James Mickens.[0] Then I fed the output to Opus 4.7 and asked it who the likely author was, and it correctly identified it as an imitation of James Mickens[1]: > Based on the stylistic fingerprints in this text, the most likely author is a pastiche/imitation of the style of several writers fused together, but if forced to identify a singl…

This is much less impressive considering how chinese models are usually copies of american models.

Re: Opus 4.7 knows the real Kelsey

#196

Post by TZubiri » Fri, May 01, 2026, 8:31 AM UTC

Stephen king once wrote and published a novel under a pseudonym to find out whether he would still be popular even if he didn't use his name.

He kept it very secret, but somehow people deduced from the writing style that this new author was the King.

Re: Opus 4.7 knows the real Kelsey

#197

Post by slabtickler » Fri, May 01, 2026, 8:33 AM UTC

deanonymization via automated stylometry is not a new idea, e.g. from 2015:

https://www.usenix.org/system/files/conference/usenixsecurit...

Re: Opus 4.7 knows the real Kelsey

#198

Post by hobofan » Fri, May 01, 2026, 8:36 AM UTC

gf000 wrote: post #192

More people should have been aware that human text contains a lot of identifiable information, and a dumb statistical model could do this a decade ago. (There were show hns with Hn user similarity analysis that used a deceptively simple model (if I remember it used like most likely word pairs only) and it was very effective. It got taken down, but the cat has always been out the bag). So your "anonymous" account coul…

For anyone interested in the details, there is a reimplementation with some explanation: https://antirez.com/news/150

Re: Opus 4.7 knows the real Kelsey

#199

Post by TZubiri » Fri, May 01, 2026, 8:40 AM UTC

littlestymaar wrote: post #159

Stylometry has existed for decades, and there's no way an LLM is stronger at that job than a specialized piece of software (it's not more realistic than expecting Opus to beat Stockfish at chess). In practice, you've never been anonymous while posting on the internet and AI isn't changing anything on that front. Or rather: if anything, AI can help you become more anonymous than before, since it can be used to hide yo…

What would be an example of such software

Re: Opus 4.7 knows the real Kelsey

#200

Post by sigmoid10 » Fri, May 01, 2026, 8:52 AM UTC

suriya-ganesh wrote: post #160

saghm wrote: post #138

Earlier quoted context omitted.

> it correctly identified it as an imitation of James Mickens How likely is it that it might take into account that it knows for sure it's not anything from Mickens from the latest training data? I'd be curious if it correctly identified a new piece from him that comes out as from him before it gets trained on it.

This is unlikely. The way model distribution works is that the model retains a lossy representation of James Micken's writing. Very likely, it cannot repeat Micken's writing verbatim. Neither can it reason about the training cutoff in this manner. It's a lossy representation

It is lossy, but it is still enough for verbatim recreations. All of Wikipedia is just 24GB of lossless compressed text and all of JK Rowling's work fits into a few MB. So these things would easily be storable verbatim in trillion parameter models. Reasoning about the training cutoff is also something that the newest models do pretty well, because you can teach them to do so after pre training using e.g. SFT. With tool use it can then even check actual current sources, which may happen without you even knowing in the normal chat apps unless you use a controlled API call.