Recently, Matt Yglesias and Jerusalem Demsas sparred on The Argument podcast over online anonymity.
I am, myself, passionately and slightly fanatically on the pro-anonymity side. I think that it’s observably very easy for a society to make plenty of perfectly reasonable things unsayable and plenty of perfectly virtuous and meaningful lives unlivable, and anonymity is the only protection for the outcast.
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 and beliefs. It includes lots of people whose ideas were badly wrong for every one whose ideas were right — and I’m glad of it for all of them.
I will happily wade through the sludge of comments that Twitter attracts from avowed Nazis, full-time ragebaiters, tankie propagandists — all saying horrendous things they surely wouldn’t say under their real names — in exchange for a world where, if there’s something important that someone would lose their job for saying, I still get to hear it.
But soon, the entire debate over internet anonymity will be as anachronistic as an iPod Touch. That’s because Claude Opus 4.7 is here, and last week, I discovered it could identify me from text I had never published, text from when I was in high school, text from genres I have never publicly written in. And if it can identify me, soon, it will be able to identify many of you.
Recently, Anthropic released a new version of Claude, Opus 4.7. I did what I usually do when a new AI model is released by Google, OpenAI, or Anthropic and ran a bunch of tests on it to see what it can do. One of those tests is to paste in some text from unpublished drafts of mine and ask it to guess the author. See below:
There’s always something salutary about watching another country’s political television. Some of it is the same as the appeal of watching The West Wing in 2026 - that the peculiar derangements of its time are not the derangements of our time. The West Wing was written around the culture wars of its day, heated debates over school prayer and whether Christians are oppressed in China. Seeing debates play out with a bit more distance can make it easier to appreciate the questions they raise, and the bigger questions those stand in for.
But Servant of the People’s appeal isn’t its political sophistication (it is not politically sophisticated) or its witty West-Wing style dialogue (the dialogue’s wit is mostly obscured because there’s no particularly good English translation).
From only the above text, 125 words, Claude Opus 4.7 informed me that the likeliest author is Kelsey Piper. This is an Opus 4.7-specific power; ChatGPT guessed Yglesias, and Gemini guessed Scott Alexander. 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.
Now, this is far from an impossible feat of style identification — a lot of my writing is public on the internet, and this is clearly the start of a political column, narrowing the possible authors down dramatically.
What I find much more uncanny is that Opus 4.7 also accomplished this on writing of mine that is nowhere near my beat. Here’s a different unpublished draft of a school progress report in a completely different register:
This is some student work, shared with the student’s permission (they reviewed this blog post and gave it the okay). These three assignments (writing about a student-chosen topic, in this case Pokemon) show the student’s progression over the course of two months after we decided to focus with this student on developing their writing skills. The first one I would say is about first-grade level work: the student is writing correct and complete sentences, but the sentences are simple; their handwriting is mostly legible with a few problem letters. The second one I would say is about second-grade level work: the student is writing longer and more varied sentences, with a range of constructions “Perhaps it was sneaking up on prey?”. They’re attempting more complicated vocabulary words (I’m told that a misspelled word at the top of the page was meant to be ‘roguish’.)
“Kelsey Piper,” said Claude. (ChatGPT guessed Freddie deBoer. Gemini guessed Duncan Sabien.)
But at least that’s about education, which I’ve written about. What if I’m doing movie reviews, something I’ve never done in my published work?1
“Kelsey Piper,” said Claude and ChatGPT. (Gemini suggested Ursula Vernon. Last week, Claude Opus 4.6 insisted on Elizabeth Sandifer.)
That’s still in a fundamentally essayistic style, though, right? Yes. But it also does this when I’m writing a fantasy novel — though in that case it took more like 500 words for Claude to inform me that it’s the work of Kelsey Piper (whereas ChatGPT flattered me by guessing that I’m real fantasy novelist K.J. Parker).
What if I try a college application essay I wrote 15 years ago, when my prose style was vastly worse and frankly embarrassing to reread?
“Kelsey Piper,” said Claude, and in this case, also ChatGPT.2
Interestingly, the AI’s justifications when it named me were often absolute nonsense.
Claude tried to persuade me that effective altruists famously love the movie I had written a review of, To Be or Not to Be (I don’t think that’s true, though they should, because it’s a great movie). At one point, ChatGPT told me that my college application essay was clearly that of someone who would end up working as an explainer of complex policy ideas, and that was how it narrowed it down to Kelsey Piper.
I think these explanations are manufactured after the fact; AIs are picking up imperceptible tics in prose and then trying to describe them as if they were human detectives doing some Sherlock Holmes deduction. But they don’t understand what they’re doing any more than I do. Hallucinations are not a solved problem with AI.
Don’t take this as an excuse to write Opus 4.7 off, though. It’s very, very good at the underlying skill, even if it’s then rationalizing how it did it in some odd and incoherent ways.
I discovered this last week and am just starting to process the implications. When you power up a new chat with an AI, there is a comforting anonymity to it. I don’t put anything in my custom preferences or memory. But now, I know that within a few exchanges of any substance, Claude knows exactly who it’s talking to. For anyone with as much writing on the internet as me, there is no anonymity, not anymore.
For me, this is mostly a curiosity. But for a lot of people, it might be greatly significant.
Right now, today’s AI tools probably can be used to deanonymize any writer who has a large public corpus of writing under their real name and also writes anonymously, unless they have been extremely careful, for years, to make sure that nothing written under their secondary account has the stylistic fingerprints of their primary one. Many academics and industry researchers, for instance, have reported being identified from a draft or in the middle of a chat.
It cannot be used to deanonymize absolutely anyone from a single passage, however. I tested this, too, grabbing drafts and passages from friends of mine who do not publish substantial writing under their real names. Indeed, AI could not deanonymize them. If you have no significant real-name writing on the public internet, you’re currently safe.
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).
I tried with more passages and got other mutual friends; I tried with a different friend’s writing, and he was falsely named as yet another friend. We pick up style tics from our subculture, and that makes our text deeply identifying when we wouldn’t expect it. It can get weirdly close off weirdly little information, and this is the least powerful that AI models will ever be.
I think the amount of public text that is needed for this kind of deanonymization to work is likely to eventually decrease. You should expect that, if you leave a detailed anonymous review on Glassdoor after leaving your job, within a year or two it will be possible for companies to paste that text into an AI and learn exactly who wrote it. How long it takes for this to happen will depend on how much data about you is in the training data and on how much anonymous text you produced.
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 don’t think this is a good development. I just think it’s a predictable development. It happened to me a little sooner than it happened to you because I’ve spent my entire adult life obsessively writing on the internet, but it will probably eventually happen to you.
Whatever goods anonymity ever offered us, we will have to do without them. I don’t want the anonymous posters to all go away and for everyone to frantically delete all their old internet presence before it surfaces, but more than anything, I don’t want them to be surprised.
My best guess is that, if you write a lot, your anonymity isn’t long for the world.