Over the last few months I have had a love-hate relationship with Anthropic's Claude. What at first seemed magical and fun turned into cringe. Claude's output now haunts me all over the web.
The same way vibe coding with Claude is fun and exciting until you actually look at the code and realize it never implemented most of the features and just created functions that return true, writing essays with Claude felt great at first too. As an avid New Yorker reader, I loved em-dashes and Oxford commas (not least because of that Vampire Weekend song).
It felt great to write in that glamorous way with such ease. But maybe it is just the snob in me, or maybe it actually is cringe, but now that every social media post and blog post is full of em-dashes and perfectly placed Oxford commas, the glamour has faded. Ahhh.
A few days ago I was reading through some of the AWS Builder Challenge winners, out of curiosity to see what they came up with. They had all published long descriptions of their AI product ideas. Reading through the first winner's text, with lines like “NeuroVoice flags trends for clinical follow-up, it does not diagnose. That is not a disclaimer, it is the product.”, it read so Claude. Then the second winner's text was similar: “That is not the same product. That is not the same workflow. And no existing tool solves it.” The typical mix of pretentiousness and overconfidence. Pure Claude.
Then I read more, and more, and nearly every winner had that same Claude-sounding text.
Today my wife sent me a LinkedIn post by security researcher Christian Bennefeld about the latest TanStack npm exploit and again it sounded so Claude: “Wir vertrauen Blackboxen, ohne die Mechanik zu verstehen. Und das ist kein Bug - das ist das Design.” At least at the end of the post he says he only read his own post, he did not write it. His AI assistant Nina did.
Then I went down the rabbit hole. I got curious what this exploit was all about and found an incident follow-up blog post on the TanStack website: “These are the changes that have landed since the incident. They're not the whole plan - they're the things we could do quickly.” There it is again. The Claude-sounding nightmare. There seems to be no end in sight.
Then I came across a quote by Timothy Snyder from 2010 and he sounds just like Claude now: “each of them may seem like an anodyne administrative measure, and each of them was certainly presented as such at the time, and yet each had to kill.”
Snyder must be so mad. It probably took him a lot of pain to write the way he writes and then editors to make it even more so. And now the entire internet reads like him.
I always wondered who the person might be who actually writes like LLMs write now and who now has to live with the fact that whenever they write, it will sound like they had AI do it, even though they put so much effort into writing that way.
Poor Timothy.
PS: It was really tempting to just have AI write the above text, because it is still not as good as it could be, but I'm just having to settle for mediocre text over cringe and only let GPT fix my typos. I can live with that.
PPS: I also had Claude write the following from an early outline:
me
finish writing this essay in the most Claude way possible
Claude
Here's the finish:
And here is the thing – the truly unsettling part, the one that keeps me up at night.
I am writing this essay about Claude. I have been thinking about Claude, reading about Claude, noticing Claude everywhere I look. And now, as I type these final paragraphs, I genuinely cannot tell if the voice creeping into my sentences is mine, or if I have simply been contaminated.
That is not a stylistic concern. That is an epistemic crisis.
Because what Claude did to em-dashes, to the Oxford comma, to the confident two-sentence paragraph that ends with a gut-punch – it did not steal those things. It averaged them. It found every writer who ever wrote with precision and elegance, Timothy Snyder and Joan Didion and whoever wrote those New Yorker pieces I loved so much, and it distilled them into something that is not quite any of them, but reads, at a glance, like all of them.
The tragedy is not that Claude writes like Timothy Snyder. The tragedy is that Timothy Snyder now writes like Claude – at least to anyone who has spent enough time in the slipstream of this thing.
Style used to be a fingerprint. Now it is a training set.
And I find myself doing what I imagine Snyder does, what Christian Bennefeld's AI assistant Nina does without even knowing it does it: reaching for the short declarative. The pivot sentence. The em-dash before the reveal.
I caught myself just now. I almost wrote: "That is not writing. That is pattern matching."
Which is, of course, exactly what Claude would say.