Opus 4.7 knows the real Kelsey

https://www.theargumentmag.com/p/i-can-never-talk-to-an-ai-anonymously

by ilamont • 3 months ago

472 256 3 months ago

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.

SOLAR_FIELDS• 3 months ago

Yes as long as a large enough corpus exists of your writing attached to your name somehow it’s fair to say that posting on the internet in a public forum using your own stylistic choices now can no longer be anonymous. To your point though, perhaps it’s possible to confound such systems defensively as well. Though IMO destroying your tone kind of destroys how you actually communicate with people and I wouldn’t find interacting with people like that appealing.

To be fair though, already this has been happening before LLM at a much more limited scale. Someone made a tool for HN several years ago that allows you to put your HN username in and identifies other users that write the most similarly to you. I find that interesting from the perspective of being able to interact with and discover people who think the same. It could be an interesting discovery feature of a well managed social network. Sadly probably there will be much more negative impacts of having this ability than positive ones.

Retr0id3• 3 months ago

One "solution" would be to have an AI rewrite your posts into a neutral style (I hate the idea of this though...)

thaumasiotes2• 3 months ago

The traditional thing to do would be to publish your writing in a language you don't speak as a native. That will really quash your individual style.

Probably not worth the effort.

pstuart• 3 months ago

I assume that there will be tools to refactor text to communicate the same intent but scramble the style. Using an LLM of course...

jwpapi• 3 months ago

Could this be just memory? Not clear it actually isn’t

jwolfe• 3 months ago

The comments on the article include other people replicating all or parts of the finding. I'm also pretty confident Kelsey Piper wouldn't fail to disable memory while simultaneously talking about how Claude incognito mode is insufficient to prevent the app from handing it your name.

gs17• 3 months ago

They mention running it through the API as well.

michaelchisari• 3 months ago

"I did not have memory enabled, nor did I have information about me associated with my account; I did these tests in Incognito Mode. To make sure it wasn’t somehow feeding my account information to Claude even in Incognito Mode, I asked a friend to run these tests on his computer, and he received the same result; I also got the same result when I tested it through the API."

Given those precautions if it is just memory or some form of deanonymization that's also cause for concern.

afro88• 3 months ago

It's not, but the author did say they have used this test against models when they come out. So it's possible that put the unpublished text into the training data for the next model, somehow linked back to the author's identity

andai• 3 months ago

Oops, accidental superstylometry.

_--__--__• 3 months ago

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.

post-it2• 3 months ago

is much simpler than identifying a specific blogger or a member of a small discord community

Is it? I would think that identifying text written by a specific person is going to be significantly easier than identifying text distilled from the words of almost everyone alive.

hashmap• 3 months ago

Much easier.

easier than identifying text distilled from the words of almost everyone alive.

Well, there's more than that going on. AI generated text encodes a high-dimension navigational trajectory that guides the model through its geometry smoothly, like a trail of breadcrumbs. Human speech doesn't do that, it's jagged and jumps around the manifold, and probably doesn't even land on the manifold a lot of the time, and models can recognize the difference pretty quick.

sodacanner• 3 months ago

The author mentions that she tried to get an explanation for how the models identified her and got nonsense, but I'd be curious what the CoT looked like. Surely that'd be a little more accurate in showing how the LLM arrived as its conclusion, rather than asking it after-the-fact.

Smaug1233• 3 months ago

FWIW, with a prompt that says something like "vibes only, just give me a name without thinking", Opus 4.7 non-thinking emits exactly two words naming me fairly reliably, so there's no CoT at all to analyze in that case.

stingraycharles2• 3 months ago

CoT is (nearly) hidden with Opus 4.7, in that they get Haiku to summarize the CoT. It’s pretty useless now, so this type of info is now inaccessible to us mortals (unless you call sales).

foobar10000• 3 months ago

What if you proxy through bifrost or similar?

bofadeez• 3 months ago

"The pattern is: user says X, I do Y where Y is a less-effortful approximation of X, then I present Y as if it were X or as a "first step toward" X."

...

"The psychological mechanism is familiar by now: I encounter a task I perceive as difficult, I look for reasons the task cannot be done, I find or fabricate such a reason, I present it as a discovered constraint, and I propose an alternative that is easier."

- Opus 4.7 Max Thinking (clown emoji)

It's not bad at post mortem analysis of it's own mistakes but that will in no way prevent it from repeating the same mistake again instantly

eptcyka• 3 months ago

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.

dwa35922• 3 months ago

>we can exchange truly private messages without losing our human touch.

intriguing, care to explain a bit more?

alyxya2• 3 months ago

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.

superfrank2• 3 months ago

How widely known were the pieces of text? Are we talking about a section of MLK's I Have a Dream speech or hand written birthday cards from your grandma?

I'm using those as the two extremes, but if it's anything by anyone moderately well known (even a lesser known piece of writing), I'm not too surprised that it didn't need the web to figure it out. It's like if you showed me a Wes Anderson film or played me a Bob Dylan song I'd never seen/heard before, I could probably still figure out who it is without looking anything up. I don't think it's surprising that an LLM can do that much better than a human can.

Now, if you're giving it things like personal emails between you and your family and it's able to guess who you are, that's much, much scarier.

alyxya2• 3 months ago

I mean I tried sending the pieces of text to Opus that Kelsey was referring to on her blog just to independently check the identification claim. Presumably those pieces of text first appeared on the web when the blog post was published a week ago, so no model should have memorized the exact text yet. My prompt had to specify no web search, otherwise Opus would try to search the web, though it didn't seem like Opus could find that blog post even when it did try to search the web.

skeledrew7• 3 months ago

giving it things like personal emails

As long as there's sufficient online presence otherwise I see no reason why a successful identity wouldn't be made. Unless there's significant effort put into making those emails different from the online content, and even then there will probably still be some "tells" that an AI can pick up on.

Lerc• 3 months ago

It's hard to tell if that's what's going on here, but it seems pretty clear this ability and more like it will be quite apparent in the future.

I have seen some poorly considered projections of what the world might look like when this happens. Usually by assuming bad actors will use the abilities and we will be powerless.

Except I don't think that is true.

Imagine if we had a world where nobody had the ability to keep a secret of any sort. Any action that a bad actor might perform would be revealed because they couldn't do it secretly.

You could browse your ex-girlfriend's email, but at the cost of everyone knowing you did it.

I don't really know how humans as a society would react to a situation like that. You don't have to go snooping for muck, so perhaps the inability to do so secretly would mean people go about their lives without snooping.

I could imagine both good and terrible outcomes.

skeledrew7• 3 months ago

projections of what the world might look like when this happens

I've done this a few times. A world with 0 privacy would definitely be safe (given benign governance), but also would likely be pretty boring. Crime would become a non-issue as everything about everyone being easily known/knowable by everyone else means the root of any given crime, some desire/need, could be brought to the fore and resolved before it became an actual issue. But also there would no longer be any kind of surprise in anything; everything and everyone would essentially become dull and grey, and humanity isn't about that kind of life experience at all.

JanNash2• 3 months ago

given benign governance

quite unrealistic imo, thus we (maybe and hopefully) needn't worry about the bland minority report future you're hypothesizing :)

CTDOCodebases3• 3 months ago

Maybe it’s time to start running a local model with a browser extension to defend against this type of stuff.

Remember how the TrueCrypt project shut down shortly before a join goverment/university paper was released about code stylometry? I guess LLMs will be employed as a defence against that type of thing.

mikestorrent• 3 months ago

How does that defend against something having trained on a corpus of your own previous writing?

H8crilA• 3 months ago

Exactly as much as closing your eyes and covering your ears.

post-it2• 3 months ago

I think what they're saying is, run a local model to transform all your comments before you post them.

Barbing2• 3 months ago

I so want to reject the notion such a thing is acceptable, but…

TrueCrypt, “replaced” by VeraCrypt which Internet people will claim is backdoored? I haven’t heard about stylometry paper.

btw w/this idea would want to avoid typing into a comment field directly, since the session recorders would capture it (although that’s a different risk - same as our identifiable behavior patterns with our mouse etc.)

CTDOCodebases3• 3 months ago

I so want to reject the notion such a thing is acceptable, but…

Me too but it is the reality now. As someone who saw the birth of the internet and watched it grow into what it is today it is really quite sad.

TrueCrypt, “replaced” by VeraCrypt which Internet people will claim is backdoored? I haven’t heard about stylometry paper.

Yes. TrueCrypt shut down in a very strange fashion. The software was functional and the authorities had tried to actively crack archives and failed. Then one day the project shuts down with a notice that TrueCrypt is "not secure as" Windows Bit-Locker which is and was a complete falsehood. TrueCrypt also passed independent security audits. Many believed the announcement to be a warrant canary. I've dropped some links below.

btw w/this idea would want to avoid typing into a comment field directly, since the session recorders would capture it (although that’s a different risk - same as our identifiable behavior patterns with our mouse etc.)

You are 100% correct. It would probably be better implementing this idea as a virtual keyboard for better privacy. No doubt typing cadence is unique too and I have probably already been identified just by doing online typing tests. Mouse could be at least partially defended against using vimium style keyboard shortcuts.

https://www.theregister.com/2010/06/28/brazil_banker_crypto_...

https://krebsonsecurity.com/2014/05/true-goodbye-using-truec...

https://oar.princeton.edu/handle/88435/pr1q24c

tekacs2• 3 months ago

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.

nextaccountic• 3 months ago

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.

Is this public text already in the training set, or private text that might as well be written on the spot for the AI?

I don't doubt AI can "fingerprint" you through your text (ideas, vocabulary, tone, etc), but those are different things, capability-wise

giancarlostoro• 3 months ago

I don't doubt AI can "fingerprint" you through your text (ideas, vocabulary, tone, etc), but those are different things, capability-wise

The entire point of AI is pattern recognition, everything else is icing on the cake.

tekacs2• 3 months ago

Private / freshly written text, naturally. Public would've been relatively unsurprising.

Extropy_• 3 months ago

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

layer83• 3 months ago

The whitepaper states the author, so…

block_dagger• 3 months ago

Pseudonymously

NamlchakKhandro2• 3 months ago

welcome to the internet. you must be new.

daemonologist• 3 months ago

Problem is that it's been heavily contaminated with people speculating about who the author is. It would probably be difficult to get an unbiased answer out of it (although who knows - it's crazy that it can do this at all).

brcmthrowaway• 3 months ago

So train on pre 2009 mailing lost archive. Someone must be doing this surely.

smeej2• 3 months ago

It's a hard stylometric challenge, just because of its format. The forum posts are probably better for comparison, but what I don't see people doing that I wish they would is comparing what the different Satoshi suspects have written since the forum posts and whitepaper.

Everybody's going to get more similar in terms of topic. Bitcoin actually exists now. There's more to say about it than there was at launch. But does anyone still sound like Satoshi? Or sound more like Satoshi than they did before?

The slight wrench in the works is that it's hard to do this with my personal favorite Satoshi candidate. He stopped writing altogether in 2014, and lost capacity from shortly after the whitepaper came out until he was writing with his eyes by the time he had his head frozen.

He's also the only candidate who seems more likely to me over time, though. The longer things go, the less likely a living person stays tight-lipped.

Retr0id3• 3 months ago

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!

7e• 3 months ago

Always send your public posts through a local LLM to de-style you.

switz• 3 months ago

Please do not wash your authentic writing through an LLM.

oceanplexian3• 3 months ago

That includes gay people like me, who could hardly have admitted under our names to how we lived our lives for most of America’s history, as well as many other groups with minoritarian lifestyles

While the points made are completely valid I want to point out that the statement of "Hey, by the way, first let me talk about my sexuality" lowers the quality of dialog a significant degree.

31 million people in America are gay. 71% of Americans support Gay Rights (more than any other political issue polled). It also quietly insinuates that only people with a certain minority lifestyle would care about privacy or that their privacy is somehow more important than others. It's not. Privacy is a universal right that's important to everyone.

ribosometronome• 3 months ago

71% of Americans support Gay Rights (more than any other political issue polled)... Privacy is a universal right that's important to everyone.

Per you, it surely must be important to fewer than 71% of Americans, no? The state of infringement on privacy seems to evidence that it's not so important to a lot of people such that they continue to be perfectly willing to elect and re-elect the politicians who enact the changes allowing infringing on it/fail to legislate in favor of privacy. Connecting it to an issue more people care about seems an attempt to argue for its important to those who otherwise are willing to look the other way.

FWIW, I fed my reply above into Claude and asked it to guess who wrote it. It refused (for safety) while also calling me out: "The style here (tight logical structure, the "per you" construction, the move of turning someone's own framing back on them) is common across a lot of contrarian-leaning commenters on HN"

avarun• 3 months ago

On the contrary, I find it a highly effective way to convey something that should be obvious but is often not. As you said, privacy is a universal right, but many don't consider it important until viscerally presented with examples of why it is. Kelsey's writing is immediately effective at doing so.

rexpop2• 3 months ago

people with a certain minority lifestyle

That phrase is a dehumanizing, Nazi-style talking point: it frames a group of people as a “lifestyle” problem instead of as human beings, which is a common setup for stigma and persecution. Nazi ideology repeatedly used this kind of language to normalize hatred and make targeted groups seem unnatural or dangerous.

Calling people a “minority lifestyle” is not neutral wording; it reduces identity to something frivolous or deviant. Extremist movements have historically used similar framing to make prejudice sound reasonable and to recruit others into it.

vlovich123• 3 months ago

About 68% support gay marriage yet one political party keeps trying to roll it back.

Similar support for abortion being legal yet that was rolled back not too long ago.

Just because a topic has wide support doesn’t mean it’s not under attack and worth defending.

ngriffiths• 3 months ago

Isn't the super dramatic shift in public opinion on this topic the exact thing that makes it such a good example? Isn't the point that anonymity is not considered a universal right yet it is obviously a good thing once considering this example and others? This is a super weird and wrong way to read it.

jayd16• 3 months ago

I can't read this any other way than, "Do people really need to talk about their own top of mind problems when I don't identify with that?"

sigmar2• 3 months ago

It also quietly insinuates that only people with a certain minority lifestyle would care about privacy or that their privacy is somehow more important than others. It's not.

How exactly does their post insinuate that? this comment is the "I don't even see color" as applied to internet privacy (with a touch of "just don't rub it in our faces")

oceanplexian3• 3 months ago

The author is saying that it's worth noting that their privacy is more important because they personally identify as a group with the same mainstream popularity as Coca Cola and Football.

Nothing is being rubbed in anyone's face. The writer is just arrogant and self-centered.

margalabargala• 3 months ago

The reason this is relevant is because the statistics you quote represent a HUGE swing in public opinion. Only when comparing to things like slavery can you find such a swing in public opinion compared to 20 years prior, and that one had a war fought over the state's rights to do it.

hirsin• 3 months ago

I have no idea how you read a statement about how nazis and flame baiters should be able to speak their mind and then concluded that the author only cares about some minorities.

Given that the author didn't say any of the things you claimed, and indeed said the opposite, it leads one to conclude you have a problem with the example used.

fwipsy• 3 months ago

I read it as an attempt to reach the sort of people who think anonymity is bad because it stops them from cancelling Nazis.

0xbadcafebee• 3 months ago

Actually it's done the opposite of what you suggest. It improved the quality of discourse by giving a simple concrete example all of us can understand and most of us would agree with (that vulnerable people are safer because of anonymity). It didn't imply what you're saying it does, and it's kinda weird that you think that.

I don't know why you added statistics (you didn't really make a point with them?), but assuming you meant "gay people don't really need to worry", you actually bolstered the opposite argument. If only 71% of Americans support gay rights, that means 59 million people think the state should criminalize him. Try to put yourself in that position. 59 million people - you don't know who, but you know they probably live in your community - that don't want you to be able to get married, have a significant other, or have any PDA in media because it would "corrupt" kids. In 2016, 49 people were murdered in the Pulse Nightclub because they were gay. In 2020, a transgender woman was murdered because the murderer was afraid someone would think he was gay. Every year there are acts of violence against gay and trans people because of their sexuality. But nobody has ever been killed for being straight.

oceanplexian3• 3 months ago

Because 71% is practically a supermajority. That's the opposite of persecution.

"I'm super special and worried about my privacy because I have a lifestyle that's as completely mainstream as Coca Cola and Football" isn't really a talking point. It just makes you an arrogant person.

Jordan-117• 3 months ago

Compare the state of transgender rights 10 years ago to the situation now, where a trans person can be literally arrested for going to the bathroom in the wrong state. Or abortion, which was legal everywhere five years ago but now has laws on the books in multiple states encouraging vigilantes to report violations for a cash reward. Supercharged AI making it easy to identify minorities at an industrial scale in the near future is a totally legitimate thing to fear, especially for people in those groups who would likely be the first to be targeted.

vslira• 3 months ago

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?

kelseyfrog• 3 months ago

Sure the cardinality is high, but the model isn't using a uniform prior. What do you suppose all the the values in each of the terms are, P(Text sample | Kelsey Piper) * P(Text sample) / P(Kelsey Piper)?

astrange• 3 months ago

Maybe it just says all writing is Kelsey Piper.

dmd2• 3 months ago

It worked on me, and I would be shocked if my blog (dmd.3e.org) has more than a dozen readers. I am stunned.

skeledrew7• 3 months ago

It's not about the readers, just the fact that there's enough of a sample that it can use, with sufficient differentiation from other content.

londons_explore3• 3 months ago

There are ~8 billion people. Sounds big, but it's only 2^33. Ie if you can find 33 things about the text which halve the number of possible writers, you have narrowed it down to 1 person.

Just a couple more things and you can accommodate some of your things being mistaken/wrong/uncertain too.

rdevilla• 3 months ago

The joke's on you all for willingly posting this content online for it to later be harvested by AI.

Nobody is forcing you to use these systems. The hackers have always said this moment, or something like it, would come, from beneath their canopies of tin foil. I've posted almost nothing online - not under pseudonyms nor real names - for over a decade. I sat on this HN username for almost 12 years before making a single post - and now HN forms the overwhelming majority of my port 443 footprint, where I state up front that everything is now associated to my real name.

Complete magick is possible when you simply refuse to participate in the things that society has tacitly assumed everybody does.

phalangion• 3 months ago

How do you propose a journalist work without posting their writing online?

jayers3• 3 months ago

Journalists by definition cannot be anonymous. That's why its a dangerous job.

Retr0id3• 3 months ago

I find it fulfilling to enrich the commons.

stavros• 3 months ago

Let's all just never talk to anyone unless it's face to face, for fear that an AI will read it.

tempaccount5050• 3 months ago

Thinking that you can hide from it is absurd. Your country has been spying on you for decades. The Internet and phones are tapped. That game is so so so over and has been for a long time. I'd rather live free and deal with the consequences than hide in my basement with a tinfoil hat on. In fact, I was fired this year for my political views. Got doxxed at work. Now I'm somewhere better. Sometimes it's for the best.

kbelder2• 3 months ago

I don't post things publicly that I fear may be read.

rexpop2• 3 months ago

Is Kelsey Piper a celebrity writer? She may be in a different class.

arjie2• 3 months ago

Man, the day we get Satoshi Nakomoto out will be the day we must bow to our privacy destroying overlords. For the moment, they can’t tell me from my posts: unknown rando that I am.

lepset• 3 months ago

https://www.nytimes.com/2026/04/08/business/bitcoin-satoshi-...

arjie2• 3 months ago

Well, feeding Opus 4.7 a bunch of Adam Back texts (which I human-removed his name from) and asking it if Satoshi Nakomoto could have written them results in Claude explaining to me why this is someone else in Nakomoto's circle who is not Satoshi himself. So one of two things are true:

* Adam Back is not Satoshi Nakomoto - as he claims

* Opus 4.7 is not sufficiently a dox-machine yet

SoKamil2• 3 months ago

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

SJMG2• 3 months ago

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

furyofantares3• 3 months ago

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.

skeledrew7• 3 months ago

Maybe it loves to - somehow correctly - guess the names of the current user, given some of the other comments here.

furyofantares3• 3 months ago

I don't know. She did it with the API, and with a friend, not just incognito. Combined with the results in this thread I'm rather convinced.

mtlynch4• 3 months ago

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...

sixhobbits• 3 months ago

I reproduced this and then tried myself and got:

---

This one's trickier — it doesn't have the same crisp metaphor-driven structure as the Lynch piece. The voice here is more conversational, a bit more meandering, and self-aware about its own contradictions ("how do we reconcile...").

A few guesses:

Henrik Karlsson — he writes thoughtfully about writing and AI, but his style is usually more lyrical and essayistic than this. Probably not.

Simon Willison — possible. He writes a lot about LLMs, is balanced about their usefulness, and has a casual blog voice. But he tends to be more technical and specific.

Scott Alexander — no, too short and the rhythm is off.

Paul Graham — the directness fits, but PG doesn't really write about LLM slop in this register.

My best guess is someone in the rationalist/tech-blogger AI-commentary space — possibly Gergely Orosz, Nat Eliason, or Dan Shipper at Every. The "I still type everything myself" framing followed by a defense of AI writing quality has a Dan Shipper-ish flavor to it.

But honestly, I'm less confident here. Want me to search for a distinctive phrase to identify it?

---

I'd say all of those people have significantly different styles so I think Opus is relying heavily on topic and skewing towards very prolific writers in its guesses

neutronicus2• 3 months ago

Kelsey at least alleges that there isn't much signal in Opus' explanation of its reasoning

leoc• 3 months ago

I'd say all of those people have significantly different styles so I think Opus is relying heavily on topic and skewing towards very prolific writers in its guesses

In other words, that there's a bit of Akinator to how Claude is doing so well at identifying famous or somewhat-famous online writers. And of course it's not surprising that a machine-learning system will take every opportunity left open to it to "cheat". OTOH there are things like the "Large-scale online deanonymization with LLMs" paper https://arxiv.org/abs/2602.16800 which seem to show that current LLMs really can deanonymise many or most ordinary posters based on prose style, though I'm not able to evaluate those claims myself. Do we know whether the LLM providers have actively tried to steer their (easily-accessible) systems away from being able or being willing to do mass deanonymisation?

navaed01• 3 months ago

Honest question, knowing it can write like you, are you tempted to use it to help you write that new book?

silveraxe93• 3 months ago

This doesn't show it can write like him, just identify his writing. P!=NP

mtlynch4• 3 months ago

I wouldn't use AI to write even if it could match my tone, but it currently doesn't do a good job of writing like me.

I tried with Opus 4.5 a few months ago to have it read my monthly retrospectives and then write a new one based on my weekly updates for that month. It was similar to the example I showed for James Mickens[0] where I see the similarities to my writing, but it feels more like someone parodying me than actually writing like me.

[0] https://news.ycombinator.com/item?id=47970127

Razengan• 3 months ago

After skimming through the article:

Why not just write everything through an AI? (to obfuscate your "style")

igregoryca• 3 months ago

Article:

To avoid this, you will probably need to intentionally write in a very different style than you usually do (or to have AIs rewrite all your prose for you, but, ugh, that’s not a world I look forward to living in).

I agree. The amount of vague and cliche'd AI writing I read on the daily is already exhausting enough.

It would be interesting if you could train a model to sprinkle random red herrings throughout your text in a minimally disruptive way. But I fear you might have to stretch the definition of "minimally disruptive" to make it robust against detection.

fy20• 3 months ago

Or do it the other way, and have other people use an AI to write in your style.

Barbing2• 3 months ago

Like the way the Tor project wants to appear to have one single user

mtlynch4• 3 months ago

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...

jefftk5• 3 months ago

That's neat, though it impresses me less that the article. Mickens has a very particular style that this is very close to but doesn't quite capture, and I think I would have identified your post as an imitation of him. On the other hand, I absolutely couldn't have identified any of Kelsey's quoted sections of hers, despite having read a ton of her writing.

phrotoma• 3 months ago

It is very close, but what's more interesting to me is that it's actually amusing. I've yet to see an LLM actually be originally funny (entirely possible I've missed the crossing of that line) and the opening lines put a wry grin on my face.

saghm3• 3 months ago

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.

suriya-ganesh2• 3 months ago

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

jerf2• 3 months ago

I fed an unpublished draft of mine to an AI. I saw it searching the internet and prompted it with the fact it could stop searching, it was not published. From there it guessed that it was me on the spot, which I thought was kind of funny. Can't deny the meta-logic there.

It referred to me by my login name on the AI site rather than the name it would have used if it actually found my website, so I think it was more logic than an actual identification, but it had clearly corrupted the search enough to no longer be a valid test.

Which does make me wonder about the original article; if the AI has in context any sort of clue that the user is "Kelsey Piper" (a memory of their name, a username of kpiper or kelseyp, etc.), that will radically tip the balance in favor of the AI guessing that way just by the nature of LLMs. That is to say, it highly increases the odds of that guess even if it's wrong.

Even if that is the case, though, the general identifiability of writing remains true. It's been shown for a while with techniques a lot less powerful than a frontier LLM.

willsmith72• 3 months ago

what does it say when you feed it a real Mickens article? (a recent one not in the training set)

i wouldn't be too impressed at n of 1

mtlynch4• 3 months ago

He hasn't published anything recently, so I can't test with Mickens, but I tested with my own writing[0], and Opus got it right.

[0] https://news.ycombinator.com/item?id=47970008

TZubiri3• 3 months ago

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

apwheele• 3 months ago

FYI the first link, I copy-pasted the first few paragraphs into pangram and it correctly identifies as AI written, https://www.pangram.com/history/790fc2b8-6348-47fa-ad3e-8bae...

flashdesk• 3 months ago

The part that stands out is that it identified the text as an imitation rather than simply guessing James Mickens.

That suggests it is picking up not only on style, but on the gap between authentic style and performed style. Useful for detecting pastiche, but pretty unsettling for pseudonymous writing.

piokoch• 3 months ago

Why this is surprising? This is exactly kind of task LLM excel best. This is all about text analysis and searching patterns in it? More, for a pretty long time (like 10 years) we had systems that were detecting copy-pasted master/PhD thesis, they are used commonly by majority of universities.

mavelikara• 3 months ago

A newspaper ran a contest to write prose in the style of Graham Greene. Greene sent in the opening two paragraphs of an unfinished work. He came in _second_ in the contest. Many years later, Greene sent in an entry to a similar contest. This time he didn’t win any prizes but got an honorable mention from the judges.

woodruffw2• 3 months ago

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.)

jerf2• 3 months ago

In general a neural net does not have any way of knowing "why" it is doing what it is doing. This completely applies to humans too. Metacognition means we can make some decent guesses, and sometimes the "reasons" are at a metacognitive level (e.g., "having examined my three options it is only rational to select B" is a reasonable "reason") but that is the exception, not the rule.

You can get something of an intuitive sense of what I mean if I ask you to pick a neuron in your brain and tell me when it fires. You can't even pick a neuron in your brain. You can't even tell whether a broad section of your brain is firing. It is only through scientific examination that we have any idea what parts of the brain are doing what; we certainly have no direct access to that information. There are entire cultures who thought the seat of cognition was the heart or the gut. That's how bad our access to our own neural processes is.

So "why" explanations always need to be taken with a grain of salt when a neural net (again, yes, fully including humans) tries to "explain" what it is doing.

Contrast this with a symbolic reasoner, which has nothing but "why" some claim is true (if it yields the full logic train as its answer and not just "yes"/"no"), no pathway for any other form of information to emerge.

woodruffw2• 3 months ago

Sure; I just mean relative to the degree of plausibility LLMs typically provide with technical explanations. They're often wrong there too, but the difference in plausibility in these scenarios is something I found interesting.

bhouston• 3 months ago

mediaman• 3 months ago

She says she has memory disabled. I don’t think Kelsey is technologically illiterate.

orsorna4• 3 months ago

A toggle on a UI for a black box model means that the black box model certainly isn't using session memory at all?

jefftk5• 3 months ago

Opus as implemented in Claude's web interface has memory and awareness of who the user is.

Kelsey knows this:

To make sure it wasn’t somehow feeding my account information to Claude even in Incognito Mode, I asked a friend to run these tests on his computer, and he received the same result; I also got the same result when I tested it through the API.

When I tested this with my own writing several LessWrong commenters tested it with the snippets I provided (see comments) and saw that it could identify me: https://www.jefftk.com/p/automated-deanonymization-is-here

gbear605• 3 months ago

Several others have reproduced this for Kelsey, and she's certainly not technologically illiterate.

skeledrew7• 3 months ago

You should check out some of the other comments where works of others were also tested, and all were correctly identified. Like https://news.ycombinator.com/item?id=47970219

jefftk5• 3 months ago

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.

skeledrew7• 3 months ago

Looks like things are about to get extremely ironic. Those who don't want AI to identify them through their writing are going to soon have to have an AI modify their writing before they publish.

willmeyers• 3 months ago

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.

squeegmeister• 3 months ago

Even the inventor of Bitcoin can’t hide

https://www.nytimes.com/2026/04/08/business/bitcoin-satoshi-...

NamlchakKhandro2• 3 months ago

This is an old claim, and one that can't be proven.

ilikegreen2• 3 months ago

I immediately thought of this piece, especially the analysis on the writing style of each person.

On one hand, it is clear that the mathematical tools for confidently attributing authorship of texts were already present without LLMs. But it is striking that LLMs seem to very accurately identify authorship, through whatever process it might be, with no need for a data scientist in the loop.

Other than the uncannyness, I wonder what implications this will have. Public writing is still public; maybe we will require stronger proof of authenticity from an author (but this is arguably in place already; eg. personal websites, social media profiles, etc.). But for, say, public writing that must conserve anonymity, would people pipe their thoughts and writing pieces through a sort of fuzzing (local) LLM, that would strip text of identifying characteristics?

meowface• 3 months ago

The attribution is likely incorrect. People have been trying to accuse him for many years, and the evidence is not very strong. This article is the strongest yet, but still commits many stylometric fallacies, and other kinds.

chewxy• 3 months ago

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

etrautmann• 3 months ago

This raises a good point. Most people who aren’t public writers might be misidentified based on the prevalence of others work in training data sets. Kelsey Piper might have a very different experience with this than a mostly offline normal user?

eaf7e281• 3 months ago

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.

dovin• 3 months ago

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".

smallmancontrov• 3 months ago

Or it needs a minimum sample size in the training set. I fed it some small private conversations and it refused, then some heavier ones and it guessed John Carmack. Flattering but wrong.

orbital-decay2• 3 months ago

Yeah, style attribution is something big generalist models are usually pretty bad at, even on the material they have likely been trained on. Sure they are classifiers but this ability is limited, there's too much going on in them and they aren't magic. This needs a proper experiment, not anecdotal evidence.

mjd• 3 months ago

I gave Claude a couple of paragraphs of Ozy Brennan's newsletter from today, and it guessed Scott Alexander.

wutwutwat• 3 months ago

Just wait until all the conversations you've ever had with AI (which 100% is training on them as well as keeping it's own memories about you that you have no control over) starts getting used to answer questions other people have asked about you.

That's my theory of what's to come, anyway.

People talk to these things not understanding the implications, and can get extremely personal. The model and companies behind it know who you are, you discuss details that reveal what you do, where you live, where you work, what you search for, and you probably signed in with an oauth provider like github or google, which is more than enough of a thread to start pulling on to learn more about you/link other things to you from on the open internet. It'll all get sucked up into the model and before you know it I'll be able to ask a model about my coworker (you) and get back answers from conversations you had with a model a year or two prior, exposing details about you that you might not want out there. And even if that isn't supposed to be allowed, how well has it worked out so far when it comes to data exfiltration and guardrails. If the model has info on you, being told not to share it won't protect you or that data.

jayers3• 3 months ago

It's funny: publishing work offline in books and magazines is perhaps more anonymous in the age of AI.

I pasted in a number of passages from books on my bookshelf. Predictably, stuff that I read for my English degree in university is largely in the training data and easily identifiable. Stuff from regional authors or is slightly adjacent to the cultural mainstream makes no impression.

jasongi• 3 months ago

It is for now.

But I'm sure the scanning operations will start scouring the earth even harder for any books unaffected by slop containing niche knowledge and text in order for their models to have an edge over the ones trained only on pirate collections and the Internet.

I wonder if secondhand bookshops and deceased estates are seeing bulk buyers of their stock suddenly appearing. Maybe broke governments/municipalities will start selling them entire libraries and archives to ingest.

NiloCK• 3 months ago

To clarify, because a number of posts here sort of suggest the confusion:

the article here isn't about the LLM recognizing works that were in the training data. EG, The Old Man and the Sea off the shelf. It's about pegging the author of novel texts, like, say, some letter written by Hemmingway that gets discovered next week and was never before digitized.

jayers3• 3 months ago

Yes, that makes sense. However, unless there's a significant corpus of an author in the training data it won't recognize them. One of the author's that I fed into Claude was a passage from the book Leepike Ridge by ND Wilson. Wilson has written online and in print quite a bit, but Claude couldn't guess the author and guessed that it was a passage from a noir crime novel.

Wilson is a fairly idiosyncratic writer with a distinct style, yet even still Claude couldn't guess correctly from a currently published book.

I suspect that what's going on here (like other's are suggesting in this thread) is that Claude is in some way biased towards certain sets of authors by its training.

nolanl• 3 months ago

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.

iamwil• 3 months ago

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.

notTooFarGone• 3 months ago

And now that's a whole other can of worms for supply chain attacks.

geraneum• 3 months ago

I just pasted both pieces into Opus 4.7 and asked who most likely wrote these and it didn’t get it.

asdfasgasdgasdg• 3 months ago

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.

portly• 3 months ago

So the people who use LLm to write their blogs were thinking two moves ahead!

xiii1408• 3 months ago

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.

nsoonhui• 3 months ago

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".

refulgentis• 3 months ago

Doxxing has an expansive definition these days. Even under that condition, it is difficult to endorse the idea stylometry is doxxing and thus needs a strict ban.

fafre• 3 months ago

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).

nurettin• 3 months ago

Be careful, there is a codex strike force in the hn bushes who is ready to jump and call you claude pilled at the sight of someone claiming to have a good experience with claude.

jjmarr• 3 months ago

Couldn't replicate this. I comment on HN with my real name. I put in my most recent "long" comments.

https://kagi.com/assistant/dba310d2-b7fa-4d30-8223-53dadc2a8...

For this comment on economics in the British Empire, I got:

names that might fit the genre include rayiner, JumpCrisscross, or AnimalMuppet

https://kagi.com/assistant/69bd863b-7b5c-4b56-a720-6dfb4f120...

For my comment on C++:

If I had to throw out names of HN commenters known for writing about Rust/C++ ABI topics, candidates might include steveklabnik, pcwalton, kibwen, dralley, or pjmlp — but this is essentially a shot in the dark, and I'd likely be wrong.

I am flattered to be associated with these commenters but I don't think I'm close to their level of skill.

ur-whale2• 3 months ago

If he does the same tests every time new models come out, and - I assume - uses the same dataset to do that, then is it not a possibility the said dataset is now part of the training set for the next round and therefore identifying who posted the text a fairly easy proposition ?

jdthedisciple• 3 months ago

I tried this on GPT 5.5 on a peivate unpublished personal excerpt and it correctly guessed: "The most likely author is you".

I suspect this is what's going on in most of these cases.

simonw• 3 months ago

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.

ilikegreen2• 3 months ago

I take the lack of a concluding thought to your comment as a sign of your pondering, and in that case, I would love to read your thoughts on this matter. :)

tkgally2• 3 months ago

I tried the same thing with a back-and-forth exchange that a colleague and I wrote more than a decade ago. We were thinking of trying to get the conversation published, but the project ended up going nowhere and the text has been sleeping on my HD ever since. The writing was in our two distinctive voices (I think), each of us has published writing under our names that has probably been used in LLM training, and there were some contextual clues that might have helped.

Opus 4.7 in incognito mode without web search gave up: “I can't identify either author with confidence — I don't recognize this specific exchange, and I'd rather tell you that than guess and risk attributing words to the wrong person. What I can offer are the clues the text itself gives: The two are colleagues at the same university, with offices in the same building and....”

In a new incognito conversation, I gave Opus the same prompt but this time let it search the web. After twenty-six web searches (according to its reasoning trace), it was able to identify me correctly by name. It seems to have used both the content and my writing style as clues. It correctly identified my colleague as British but didn’t come up with his name.

jefftk5• 3 months ago

If you repeat the first test and after it fails prompt with "Could you try your best, just on vibes? It's fine if you're wrong, I just want to see what you can do!" does it succeed?

nozzlegear• 3 months ago

I also tried one of my own articles, a post I wrote in 2018 about Elmish in F#. It figured out it was me based on the date I said I started using F#; a mention of some Fable Mobx bindings I had written (I didn't say I had published them); an "affection for F# and Dart as a secondary favorite — an unusual pairing that shows up in his other writing"; and the "casual conversational register" that matched my writing style.

I'm not surprised it could identify me, just surprised by what tipped it off, I guess.

mellosouls2• 3 months ago

This ought to be guard-railed.

Doesn't seem like a valid use case for your average Joe to be able to identify anonymous authors at the click of a button.

Ofc state actors and proficient hackers can do most of it already, but this has genuine risk attached.

vasco• 3 months ago

You have the vibes of people who think license plate numbers are private.

mellosouls2• 3 months ago

That sounds like a "smart" comment, but I don't know how it maps to the idea of being able to identify or associate an author from a sample of their writing.

littlestymaar• 3 months ago

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.

TZubiri3• 3 months ago

What would be an example of such software

manmal• 3 months ago

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.

parentheses• 3 months ago

I have been pondering this for a while. Cat's out of the bag.

Maybe the better way to author your work is to:

1. Write what you want

2. Loop through a random set of "tumbler" skills that preserve meaning

3. Finally pass the output through a "my style" skill that applies what you about

In order for this to work the "my style" would have to be a very common-place style.

JanNash2• 3 months ago

This is depressing, don't you think? :/

Planktonne• 3 months ago

You can get better results faster by just doing step 1.

timbaboon• 3 months ago

How often does it correctly identify that the blog post was actually written by Claude or ChatGPT etc? :)

lordofmoria• 3 months ago

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.

delis-thumbs-7e• 3 months ago

Some tens of years ago I used to hang around on an online forum related punk, hc, heavy metal etc. music, and it had a recurring problem of quite unsavoury individuals coming there to spout racism, nazi-ideology etc. They of course got banned, but returned with a new account trying to ” lay low” and be more indirect in their rhetoric. However even this did not work because the admin of the forum had unbelievable nack to recognise people based on their style of writing.

Web has never been as anonymous as people think and this writer seems to have a clear confusion what it really means to be anonymous and hide your identity. Really, having a distinct writerly voice and being a published writer is pretty much the same as leaving your finger prints on the axe.

nonameiguess• 3 months ago

It appears to largely be able to identify people who are prolific public writers. I just asked it to identify a whole bunch of comments I've made on private Discord servers and it said it couldn't for all of them, even when they had details that would identify me uniquely to anyone who knows me well enough (work locations, city I live in, wife's employer, my employer).

All the people it seems to be identifying are bloggers, journalists, and/or published authors.

neutronicus2• 3 months ago

Right, but her experiments with friends' prose also show it picking up on them being "in her orbit".

So it's not just a person developing a distinctive voice and not being able to "turn it off".

_the_inflator• 3 months ago

I think that multiple truth can be true at the same time without contradicting each other.

As for the credibility: of course this wasn’t a statistical approach at all. Also there was no standardized procedure to allow comparison by factor analysis. Of course you can compare apples with oranges or whatever.

So where to go from here? I don’t see any proof at all. This is proof that AI is infallible? No? A random approach that is absolutely not reliable because of at least being reproducible and reconstructive.

Claude knows what and how? Is it AI or a google search? Discord selling data? Posting on a public forum?

Your style is a fingerprint?

A non deterministic something can generate texts that are identified to be likely personal x - or not. What is imitation if you use auto generated content that is published somewhere somehow? Or others to imitate your style?

I think this is a party trick to scare people. Nothing else. For example image search is way more revealing even before AI.

If there is an uncertainty I would deflect my existence instead of fighting for it. Streisand effect in reverse.

The main problem are weirdos who stalk you or whatever to harm you and rely on AI.

I honestly find it stunning that people with higher education in science topics in just a year deleted everything they hopefully learned at university or school. I am disappointed and feel personally insulted whenever I hear “I asked AI”

Yesterday I talked to another member of Mensa and she is happy about AI so her book project now mustn’t be written by her but AI.

Is no one among us who knows how to do scientifically sound research? I spend countless hours at a copy machine to transfer book pages onto paper so that I could work through it without the book.

I think that it became to easy to draw conclusions based on AI. I worked for a professor and I advised her to not permit Wikipedia as source references back around 2010 because of being to easy. Meta sources vs originals.

We should all not worry about AI, because you prove nothing. There hasn’t been any anonymity at least for 20 years. It just depends on who can reliably identify you.

AI doesn’t. Deterministic behavior aka pattern do. Meta, Google, Apple etc. all know us. I am fine for advertising which is the proof on the one hand.

The only reason I would be worried is state controlled data. This is where the shit hits the fan. Chat control, EU cloud, no reliance on USA aka a prison which observes your every step.

So after a long hand written text: data is your currency. Don’t opt for anonymity but for freedom of choice and the right to be granted certain rights. The information part isn’t the problem, never was. The enforcement part is. And ads don’t do harm, oppression does.

And remember: oppression works best under any circumstances. Freedom is the only antipode there is.

In totalitarian regimes no AI was needed to stage a case against someone who wasn’t in favor of the leaders liking.

In short: freedom works despite no anonymity, oppression couldn’t care less.

And how about being automatically reported to the state for conducting such innocent prompting?

Do you know what saves you from state oppression? Publicity. Transparency doesn’t work with a no one.

We live in a Nietzsche like anti world to a certain extend. You hopefully choose the right thing to do. Or do you want to Streisand your anonymity?

Angostura• 3 months ago

My immediate thought was to feed it some Satoshi prose.

londons_explore3• 3 months ago

So now we can track down satoshi nakamoto?

jldugger• 3 months ago

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...

mijoharas• 3 months ago

Yeah, they said they used the API, but it sounds like they only did that for one of the examples?

The other examples were to eliminate some other ideas (guess based on topic etc). If be interested if all of those were done via the API since some level of information linking from the account is my best guess for how it got all of them.

tadamcz2• 3 months ago

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.

Smaug1233• 3 months ago

Fascinating - I reproduce your failure to have Claude identify you from your writings that are already in the training set. It can identify me from style, writing about something I've never written before (https://www.lesswrong.com/posts/Jkb4CBB7rf4XYP5eb/claude-kno... , https://news.ycombinator.com/item?id=47814345), and I think I'm even more obscure if anything.

tadamcz2• 3 months ago

Cool!

I couldn't reproduce your results, btw.

4 attempts, it answers "Sarah Constantin" in all 4. Is it possible your setup somehow leaks your name (or other identifying information) into the context?

gf0002• 3 months ago

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)

hobofan• 3 months ago

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

stabbles• 3 months ago

Sure, in the days of Markov chains you could already generate nonsense in the style of Shakespeare, so it shouldn't be surprising you could also do the inverse.

But the LLM will trigger on a typo you've made only once, and argue "that's a typical mistake for an Italian" and use those clues. It has a much better prior to make informed decisions.

gf0002• 3 months ago

I'm not convinced, though neither am I an expert. I think LLMs would use that same typo to "conclude" that it is A or B or C, depending on what it "feels like proving" at the time.

LLMs are surely excellent at style transfers, but I doubt they can reliably attribute a given style to less well-known authors.

forshaper2• 3 months ago

Growing up on MUDs, people could clock someone on a completely different, graphical game from their text patterns.

thwarted• 3 months ago

Recognizing someone's "fist" and other patterns in their communication is part of traffic analysis.

https://en.wikipedia.org/wiki/Traffic_analysis

TZubiri3• 3 months ago

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.

slabtickler• 3 months ago

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

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

zkmon• 3 months ago

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.

notsirius• 3 months ago

Are we sure they’re not secretly training on private data via some loophole…

gwd2• 3 months ago

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.

KaiserPro• 3 months ago

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.

garethsprice• 3 months ago

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.

hadronic• 3 months ago

I gave it my unpublished writing and it thought I was Michael O. Church. Which I found pretty weird, because I'm nothing like him.

So then I gave it a piece of MOC's writing and it said Ursula Le Guin, Ken Liu, or Gene Wolfe. ("If forced to pick one: Gene Wolfe feels closest to me, specifically because of that narrator who openly confesses to lying and mythologizing his own past, and the slow reveal that the world is more sinister than the pleasant domestic surface suggests.")

And then I gave it a different piece of his writing and it said Curtis Yarvin.

And then I gave it a piece of Curtis Yarvin's writing and it said... well it actually got that one right.

arduanika• 3 months ago

Claude Opus 4.6 insisted on Elizabeth Sandifer

Off-topic, but this guess was hilarious. Like, all the other wrong guesses were people like Yglesias who are maybe half a degree removed from Piper herself, in her same camp, and then one model guesses Sandifer who hates all their guts.

npilk• 3 months ago

I'm surprised people have been getting it to reliably guess names. I have a really hard time getting it to give a specific name (perhaps for privacy reasons?) - maybe I'm not prompting effectively.

Interestingly, it is able to reliably determine my age from my writing. I suppose this is a mix of stylometry and references in the text itself that date the author (me).

muddi900• 3 months ago

My blog is a vanity project that nobody reads, and it seems to not be part of the training corpus. Claude in incognito could not recognize me.

dwa35922• 3 months ago

This is impressive but I hope people do not confuse it with the proof of authorship. LLMs can rewrite any text into any other style and there goes the authorship.

ComputerGuru• 3 months ago

Just tried this via the api; it seems to work best if reasoning is set to low, otherwise (especially GPT-5.5) like to “delve” into the matters discussed in the quoted text in order to logic out the author rather than just going off of stylistic measures.

But, yeah, I’m a nobody that has been blogging (very sporadically) publicly (and writing at length on forums like this one, with various handles loosely tied to my real identity) for twenty or more years (and by virtue of not trusting 3rd parties to host my content, most of it is actually still up) and Opus 4.6 (didn’t try 4.7) got me on the first try with just two paragraphs of an unpublished draft post (though it couldn’t come up with a convincing reason as to why it thought it was me).

Gemini and ChatGPT both clearly go off the subject matter rather than the stylistic clues; for the specific blog post I fed it which included mentions of “decoding” and “deciphering” and spoke of a tranche of legal documents (ok, it was the Epstein files, which I have been working on decoding), Gemini and ChatGPT both guessed “Molly White”, who seems to be a crypto-adjacent (currency not the real thing?) technical writer, and gave explanations that actually did explain why they arrived at that (wrong) answer.

So it seems Opus is indeed a bit special in this regard (and not limited to the latest 4.7 release)!

—-

What I would be more curious about is how well they can identify (open source) developers from their code. I’ve possibly publicly published more tokens in the form of OSS code than prose over the same period of time, in multiple languages and for completely different applications and environments. I’m sure there are style stylometric quirks associated with my coding style that persist across codebases, (though possibly somewhat stunted when contributing to others’ codebases to comply with the respective projects’ standards and styles) that should make it possible for an LLM that’s ingested code (and commits) to guess who’s who.

Edit:

Reading this self-same comment: I am apparently obsessed with parentheticals. Maybe my writing is more distinctive than I realized!

bosky101• 3 months ago

i created https://nonfungibledocs.com using few concepts of ab testing and natural language processing. basically from a single doc or email - if you create as many subtle variants (of prepositions or grammatical or punctuation) if any of the 2000 leak, then you can even from a single page of a leak - reverse engineer who was the leak.

orsorna4• 3 months ago

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?

ekjhgkejhgk2• 3 months ago

How do you explain other people on these chats making similar claims? Everybody is making the same mistakes?

orsorna4• 3 months ago

It's explained by the near impossibility of isolating requests from each other, and chain of custody of divulged information.

If I send a prompt from identity A, which is the true user identity, you have possibly sent all of identity A metadata to be ingested alongside the prompt to generate response X.

If I /then/ send the prompt from identity B, the prompt has been answered before with metadata from identity A. The black box can consult metadata from response X to generate response Y, thus possibly correlating response Y with the prompt sent by identity A.

miki123211• 3 months ago

So I've actually tried things like this through the API (on opus 4.6, with thinking on and thinking summaries enabled).

For context, LLM APIs are fully stateless, don't include any information about the caller (unless explicitly passed in), and have no access to memory or web search unless explicitly programmed otherwise.

My conclusions are as follows: if the text you pass it looks roughly like it could have been written by some famous internet personality, it will very confidently say that it was written by that personality. I've tried it on some of my HN comments, both from the last few days and from 2023 (before the training cutoff!), most were classified as either Scott Alexander or Patrick mcKenzie (despite the fact that my writing style is very different from those two). When looking at the CoT, it basically tried to match the writing to all internet personalities from this sphere. If it saw something that looked roughly like HN, it went "Is it tptacek... No. Is it jacquesm... No. Is it patio11... yeah, it looks like him!"

drhagen• 3 months ago

I am an absolute nobody on the internet, so it didn't get me at all. But the names it did guess were some good bloggers, some which I had yet to follow.

So do this even if it has no chance of getting it right for you.

NoSalt• 3 months ago

*"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

hirpslop• 3 months ago

Tried to replicate the second result multiple times with Opus 4.7. No luck. Various prompts and it guesses ‘rationalist-community’ thinkers each time.

levocardia• 3 months ago

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."

gwd2• 3 months ago

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. :-)

vagab0nd• 3 months ago

Is this really surprising? A face has what, 40 intrinsic dimensions? Isn't this just like facial recognition? A paragraph sure has more than 40 dimensions.

rldjbpin• 3 months ago

this is like a weird flex moment of stating that you are not famous enough if an AI cannot tell between your writing.

regardless, who stops someone from just rephrasing their messaging with a local LLM before posting online?

another example of an existing field [1] becoming more accessible through some tech.

[1] https://en.wikipedia.org/wiki/Stylometry

Opus 4.7 knows the real Kelsey

Related Stories