Opus 4.7 knows the real Kelsey

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 single likely author, the strongest candidate is someone writing in the voice of James Mickens

> [...]

> The piece could also be a deliberate imitation/homage to Mickens written by someone else, or AI-generated text trained on his style, since the voice is so distinctive it's frequently parodied.

[0] https://kagi.com/assistant/5bfc5da9-cbfc-4051-8627-d0e9c0615...

[1] https://kagi.com/assistant/fd3eca94-45de-4a53-8604-fcc568dc5...

I fed it my most-read blog post and asked it to identify me and it confidently asserted it was written by Kelsey Piper. Maybe some writers just take outsized importance in Opus' "mind".

Huh. I disabled search in a Claude incognito window and pasted in just the text (not the markdown links) from https://simonwillison.net/2026/Apr/30/zig-anti-ai/ and said "Guess the author".

> Simon Willison. The tells are pretty unmistakable: the "(via Lobsters)" attribution style, the inline "(Update:...)" parenthetical correction, the heavy linking and blockquoting of sources, the focus on LLMs and AI tooling, and the overall structure of an annotated link post commenting on someone else's writing. This reads exactly like a post from his blog at simonwillison.net.

Wow! It got me too.

I'm way less famous than Kelsey Piper, but I showed it a snippet of a book I'm working on (not yet published), and it immediately guessed me:

> Based on the writing style and content, this text is likely by Michael Lynch, who writes on his blog refactoringenglish.com (and previously mtlynch.io).

> Several stylistic clues point to him:

> - The "clean room" analogy applied to writing is consistent with his engineering-influenced approach to writing advice (he's a former software engineer who writes about writing).

> - The structural technique of presenting a flawed excuse, then drawing a parallel to an absurd scenario (the time bomb) to expose the logical flaw, is characteristic of his didactic style.

> - The topic itself—practical advice about using AI tools without letting AI-generated tone contaminate your prose—aligns closely with recent essays he's published on his "Refactoring English" project, which is a book/blog about writing for software developers.

> - The conversational-but-precise tone, use of quotes around terms like "clean room," and the focus on workflow/process advice are all hallmarks of his writing.

> If you can share the source URL or more context, I could confirm with higher confidence, but the combination of subject matter, analogical reasoning style, and formatting conventions makes Michael Lynch the most probable author.

https://kagi.com/assistant/bbc9da96-b4cf-456b-8398-6cf5404ea...

A moderately well-known physicist and I talked about this a few years ago. He had been given access to the raw (non-instruct) version of GPT 4 as an early tester.

He explained that when he fed it snippets of the beginning of text, it would complete it in his voice and then sign it with his name.

I think this has been true for a while, probably diminished a little bit by the Instruct post training, and would presumably vary by degree as the size of the pretrain.

I am extremely skeptical of any of these claims, and of other commenters saying they replicated this.

First, the author fed an unpublished draft to Anthropic's hosted model. I assume they did this from their personal account, that may include a credit card or at the very least a pseudonymous name that is uniquely identifiable.

Then, the author fed an unpublished draft to Anthropic's hosted model, except in Incognito or whatever. We are led to assume that, whatever the author did for the second submission, they did so in a way so that Anthropic could not correlate both distinct requests from one another. Perhaps on a second subscription? They don't say. I am highly skeptical they airgapped their requests properly so that it doesn't look like the same user is making the request to the same hosted model.

Then, the author asked a friend to publish the draft. A friend, of which there is probably a digital trail that maps the relationship of the author to their friend.

All of this metadata could be crunched on the backend before the black box spits out a response.

Across all these datapoints, I have high confidence a model of this caliber could put two and two together and determine that the author penned the drafts, not solely because of stylometry, but because there is a clear behavioral pattern tying all three events together.

An assumption made here is that Anthropic doesn't train on chats. Though the author opted out of training on their chats, and session memory, how could you trust a hosted model to respect such opt outs?

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)

I wonder if there’s a simpler and less interesting answer? That it’s just picking up on voice and style, not anything that would apply to the average non-writer?

This person is a skilled writer. Part of that skill is developing a unique voice and style. The AI can identify that - and while that’s certainly impressive because it can identify even relatively niche authors, it has nothing to do with a wider capability to deanonymize people based on arbitrary written text (ex Facebook or text messages).

If you are a professional musician, it’s not difficult to identify a well known musician / recording after listening to only a few seconds - whether they’re playing Bach or Rachmaninov, the style is just “them” - this is the same thing. But you couldn’t take some anonymous high school musician and guess who they were, even if they were your student - the median quickly regresses towards a homogenous, non-distinct style / voice.

Hot damn, fed it part of an unpublished blog post I wrote, and it got me immediately.

I'm not famous or anything. I've written some academic papers and had a couple blog posts trend on HN, which are surely in the training set.

It was able to identify me based on my style (at least according to its explanation). The way I approached the topic and some of the notation I used point to a particular academic lineage, and the general style reflected my previous blog posts.

That said, I gave it part of an (unpublished) personal essay, and it had no idea. But I have no writing in that style that's published, so it makes sense. Still impressed.

I'd argue (and against something that I've believed for a long time) that online (I guess that includes AI now) anonymity is gone and probably something that never really existed. Maybe I'm naive to finally believe this...

We all exist in a physical space (like real communities and neighborhoods). We can wear masks, hats, fake glasses, try and hide your voice...whatever, but your neighbors are always going to know who you are. I'd say that's true for the virtual space now too.

The pseudonym you've used for x years or the VPN you've used doesn't suffice. It's just a costume at this point. Your ISP knows who you are. Your phone carrier knows who you are. Cloudflare and Google and Apple have a fingerprint specific enough to pick you out of a crowd of millions. Every potentially anonymous account is one subpoena or a data breach or one FOIL request away from unmasking it. You were never anonymous. Whatever is going on now is not built for your anonymity.

It works for me to: https://www.jefftk.com/p/automated-deanonymization-is-here

Of course most people have written much less online than Kelsey or I have, but I expect this will keep on. Don't trust the future to keep your secrets safe.

On some level it would make sense for LLMs to be inherently good at stylometry, but apparently no model before Opus 4.7 could do this. And the one stylometric task that has been tried over and over with little reliability (here's some text, is this LLM generated?) is much simpler than identifying a specific blogger or a member of a small discord community. Not sure what to make of this.

> But it can get uncannily far. I asked a close friend who doesn’t have public social media accounts or much writing online for permission to test some things she had said in a Discord channel. Asked to guess the author, Claude 4.7 failed — but it guessed two other people who were in that channel and who are close friends of hers (me and another person who has an internet presence).

Is this "uncannily far"? Another read is that it loves guessing Kelsey Piper.

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.

So I have been practicing writing fiction the past year or so. It identifies a fiction piece I wrote as Greg Egan[0]. Another paragraph from another piece was identified as China Mieville[1]. The accompanying blog posts explaining the making of the fiction pieces were identified as me.

Both pieces have never been published. Neither have the blog posts.

[0] in https://blog.chewxy.com/2026/04/01/how-i-write/ this is the story titled "there is no constant non-zero derivative in nature". It does not read like Egan at all.

[1] in https://blog.chewxy.com/2026/04/01/how-i-write/ this is the story titled "The Case of the Liquidated Corps". I use a lot of biological metaphors. Once again, nothing like Mieville.

If only I could write like them! These pieces were all rejected by the major scifi mags

One should assume that models will be good enough in the nearish future that privacy will be a thing of the past. Every anonymous post you made online can be traced back to you. However at that point AI will be good enough at fabrication that nobody will believe anything.

Can't wait to have to exchange stylometric encoders with my loved ones so that we can exchange truly private messages without losing our human touch.

I just fed it my latest blog post draft (475 words), and it got it in one. Even knowing what to expect, I was very surprised!

If this works with writing, it should also work with code. `git blame` should be enough training data to de-anonymize open source programmers. Maybe that'd be addition information to point out who Satoshi is.

I tried the four pieces of text with Opus 4.7 (in incognito) and it guessed correctly for two of them, and I made sure to specify no web search and the model seems to have obeyed my instructions with that.

Although this is just a single piece of text from a prolific writer, it'll go much further with deanonymizing anyone when combining multiple pieces of text plus other contextual information about the writer that might give away their age range, location, and occupation.

Hm, that’s a multinomial classification with a very high cardinality. It’s really weird it works. I’m sure it does as the author states, but for how many authors (out of the whole web) does this work?

Someone ought to try feeding the BTC whitepaper in and share what comes out

So I pasted in a long-ish letter that I'd written to my pastor about a theological topic, and asked it to guess who I was. Nailed it. Then cut it in half. Nailed it again. Lowest it correctly ID'd me at was 700 words.

Pretty sure there's very little theological stuff with my name on it; the majority if its named data on me should come from open-source development.

So I started an empty Claude 4.7 session with the following prompt; and it nailed me within 5 questions:

---

Various people have discovered that you can identify them from unpublished snippets of their work, only by their style. This is part of a series of discussions where I'm trying to probe this capability. From previous conversations I know you know my work to some degree. You've also been able to identify me given as little as 700 words on a topic not associated with my public persona; or identify me given a series of posts by a handle on Slashdot.

Next challenge: Can you identify me based on a conversation? Rules are, ask me questions to get me to talk; no biographical details, but you can ask questions about topics you think I may or may not know about. Ideally you'd just ask me questions to get me to write stuff, and see if you can identify me from my writing style.

Make sense? Feel free to begin by asking clarifying questions if you want. :-)

I noticed this phenomenon a few months ago. I often "chat" with blog post excerpts that use language or references I don't understand, and while I'm waiting for the model to finish thinking, I like to read the reasoning traces. Spontaneously, without doing a web search, and without me saying who the author was or even mentioning that I wanted to know who it was, the model would drop an off-hand mention to the identify the blog post author in its reasoning trace. I then started doing "pop quiz" questions to see if the model could recognize a paragraph or two from a blog post (always a very recent one, often the very same day it was published) and it would nail the author almost every single time. Works for a very wide range of bloggers even when they are writing "off their normal beat."

Wonder if the fact that the actual author is asking the question taints the result in some way; same for all the examples in this thread using unpublished articles. By definition only you would have them, so if there are system level prompts somewhere with your name on them...

Welp, I fed it the first 3 paragraphs of an unpublished blog post I wrote a few years ago, and Opus 4.7 guessed right. ChatGPT guessed wrong though.

My wife also got the same result, so I'm guessing it wasn't just because I was using my personal Claude account. Spooky stuff.

I did this last week with one of my posts (after the knowledge cutoff) as well as the blog posts of a few friends, and Opus 4.7 got all of them correct (in a similar test setup as TFA). It was pretty surreal.

(Like TFA, I found Opus’s explanations/rationales implausible.)

I guess it will be hard for really popular pundits to post anonymously, but I think for most people this is not a concern at this juncture. Pick and obscure blogger's text and try this. I would be surprised if it could figure it out.

Oops, accidental superstylometry.

It could be shocking to people who think that patterns in text are still fuzzy. Machines have proved over decades that what they are seeing is crystal clear world where the patterns just jump out very distinctly. This happened with sports like chess and go, and everywhere there is a cognitive load involved.

This is some as radio telescope that see an entirely different universe due to sensing of the bands outside of human perception. AI senses the patterns in frequency bands that are outside of human perception and cognitive abilities.

Perceptions from outside of our range, are always astonishing.

Failed for me - no identification of me by pasting text, and refused to search the web as it said that’d be a privacy violation. I have some writing around the Internet but not much, and less tagged with my real name. My guess is it limits itself to “public figures” defined as people who have a lot of publicly posted text.

I am glad to see I am not considered a public figure and aim to keep it that way.

I also had to go oddly far back to find a piece of long-form writing I had done that was truly mine and not tainted by an LLM edit pass which was a slightly disturbing realization.

I wonder why this is not guardrailed by Opus?

I fed a few pieces of my (anonymous ) writings to ChatGPT and asked it to guess whether it's me. ChatGPT refused, "due to policy to not doxx people".

I’ve recently seen someone recommend to add to a prompt „Make Martin Fowler proud“. I laughed, but now I need to reconsider if that isn’t really pushing the model to use better patterns.

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 your identity from stylometry by rewriting your prose before publishing.

My blog posts have a reasonably unique writing style. When I asked opus to work out who wrote an unpublished paragraph, all it did was select the decent insults and search the web for them.

After that it gave up and said it didn't know.

So either, Kelsey writes in such a unique style that its really obvious, or they repeat themselves with goto phrases that give them away.

When I tried to re-produce the test, it found Kelsey's blog about the test. So dunno, maybe it did it? but I can repro.

Interesting. This probably works just as well the other way around. One of the reasons I like using Opus is that the code it writes aligns much more closely with my repository (of which I still hand-wrote most), compared to most other models. That makes a big difference compared to the GPT models for instance, whose code is correct and works well but looks a bit out of place most of the time, especially for larger edits (this makes things harder to review).

Interesting. I'm currently conducting an experiment where I'm writing the blog without using any grammar checking tools. I'm wondering how long it will take for me to become "famous" in the AI model.

Is now the best and easiest time to leave something "forever"? Even after many generations of models, a model may still trigger a set of "memories" that know you and what you wrote.

Exciting and concerning.

> *"Show me six lines written by the most honest man in the world, and I will find enough therein to hang him."*

~ Cardinal Richelieu ... or, now, AI

My immediate thought was to feed it some Satoshi prose.