2,704 words about the Claybot

Why I fed all my writing into an LLM, and what I learned from it.

I recently fed 200,000 words of my writing into a large language model. I spent days putting the corpus of work together and an entire lifetime creating it, and within 15 minutes it was telling me things like:

You possess a unique ability to reference Noam Chomsky, Michael Mann, and Jorge Luis Borges in the same breath as whippits, trashy pop-up ads, and bodily fluids

And:

You often describe the abstract experiences of the internet or corporate life using unsettling biological imagery, such as “viscous residue,” “purple pustules,” and “blackened slime”

And, a personal favorite:

You use your capacity for high-level analysis to “rip on absolutely everyone” as a means to survive the “reverberating infinite blankness” of the universe”

Why did I do this? The same reason I do anything: screaming ontological terror at the precarity of modern existence. Like everyone else, I am worried about artificial intelligence. As a writer, I’ve already seen it decimate vast swaths of labor that once paid my bills: SEO and SEM work, research reports, social copy, and don’t get me started on journalism, which was a lightly compensated hobby to begin with. One of the ways writers have combatted this is by rebelling as a cohort against AI-generated writing, and filing, sometimes en masse, lawsuits about the pilfered content used to train the AIs.

But it’s also hard not to see how much better this stuff is getting. Put aside the eye-popping stories about Anthropic’s latest model uncovering decades-old security vulnerabilities. If you use these tools at all – which, anecdotally, most people who work on the internet at this point do – then you know all the crazy things the latest models do. It can disambiguate a meal’s nutritional content based on crappy snapshot. It can spin up a website for you that looks good pretty in like an hour. I use it as a sort of junior employee almost every day to wrangle recordings and handwritten notes into coherent documents and next steps. AI writing is still bad, in other words, but what if someone tried to make it good? How close could it get?

Recent articles in the New Yorker, the New York Times, and the Atlantic have earnestly grappled with this, to varying ends. I am, if nothing else, a competitive writer, and I found myself stymied by the thought of publishing much of anything without at least engaging this threat. Not to my livelihood, but to something deeper – my “voice,” that blend of vocabulary, rhythm, diction, disposition, structure, and aura that makes a piece of writing feel distinctly by someone.

It felt unethical to make it replicate the voice of, say, Cormac McCarthy, who I presume would haunt anyone who undertook so a base a conjuring. But I could fairly ethically train it on my own writing, which there is certainly plenty of out there. I started publishing online over 20 years ago, and I wrote hard: long, passionate blogs in the early days, lots of literary nonfiction and personal essays throughout, a good bit of pop-culture reportage, and so on, all dripping with voice in a manner that some editors found frankly distasteful. It wasn’t all good, but some of it was, and even when it wasn’t, it was unquestionably by me, which made it a great test case for the technology’s ability to adopt a writer’s voice rather than the cheery, emoji-spackled UX-ese it normally uses.

Hence: the 200,000 words. I first trolled the archives of my various career stopping points to collate a Norton Anthology of Myself – the outrageously overwrought Weezer review from ‘05, that one spacey Ghostface blurb, the big Netrunner exegesis, the LARB stuff from a few years ago, plus a whole lot more. Like everyone else, I wrote a novel during lockdown, so I threw that in – it’s got a ton of good sentences and it’s pretty indicative of my whole “thing” even if, as a narrative, it never cohered. Once all of that was in a single sheet with labels (publication + year) I started shopping the process around various LLMs to get their analysis.

They buttered me up, of course. That’s what these things do. But then they started to generate strange, hallucinatory copy that wasn’t necessarily about anything, sometimes transmuting phrases verbatim from my corpus, other times circling a single anecdote from the novel in a cold, airless hell-state. This all felt like mere verification of the thing every writer wants to see: that an LLM can’t possibly create writing with the spark of truth, of lived experience. So I asked it to write an essay about the only thing I knew it could speak to authentically: that is, the experience of being an LLM trained on my writing specifically. This generated something usable, substantive, and relatively new, I felt, so I iterated. I spent a good while attempting to create a system-level prompt that might automate my voice, an exercise I would recommend to any writer seeking absolute souldeath. One LLM eventually described me as “holding court, two drinks in,” a description that helped the ensuing test copy, so we dialed it up – three drinks, five – until it got too drunk and admitted, midway through an essay, “I am not good.” This was funny, but a bad model of my voice, which would never get so emo.

Ultimately, the prompt that worked was no prompt. I popped the essay idea into the chat where I’d been developing the prompts and what it spit out worked top to bottom: well-structured, eerily close to my own voice, substantive in its analysis of its own processes, and, in the final set of sentences, shockingly strange. I knew I had it.

///

You better believe I burned some tokens after this breakthrough trying, and failing, to create a five-essay series that graded imperceptibly into speculative fiction: the AI consumed my body, populated the world, the airless hell-state came to life, etc. There’s a reason my novel didn’t work. I pretty much want every plot to be Neon Genesis Evangelion.

But attempting to create the NGEGPT saga did generate a lot of new, lived experience within that single instance of the chatbot. I thought there may be a new narrative within those trials that it could articulate, so I prompted it one more time, asking it to describe all of my failed attempts to follow up the first essay, and the result was part two. I asked it to surprise me at the end, and it did, again: It told me to quit. It told me to stop making AI essays and to go write something else. It told me to keep writing, far from the Claybot.

Was this flattery? Of course. It was also a little bit paralyzing. The response to the first two essays – which I published here because anywhere else seemed unethical – was as bifurcated as you would expect. A handful of readers signal-boosted the heck out of them. Old friends texted me, somewhat moved. Others were less charitable. Former colleagues quietly cited them as evidence that I had always seemed untrustworthy. Some of the people closest to me were genuinely offended that I would even engage with the technology – it signaled some new technocratic complicity in me, something not only not particularly Clay but not particularly human.

I admit that part of the initiative for this project was churlish: an attempt to provoke, to willingly feed my life’s work to the fire, when so many writers are flailing in the opposite direction. I’ve always written for attention, an audience, a response; I was, and am, annoying. But I do find myself sincerely changed by the experience of creating the Claybot and publishing its output. I knew I wanted to earnestly attempt to create something good out of it, to get as close to the fire as possible, but I didn’t expect to enjoy the whole humiliation ritual as much as I did. The increasing quality of generative AI is both writers’ block and inspiration in one: it sucks everything in, demanding attention, content, kindling. For some people, this is proof positive of the problem. But all I have ever written about, really, is the internet and the way it changes culture. I think the central image of the entire corpus I fed into the LLMs is a feedback loop, growing increasingly discursive and strange with each iteration. The fuzzy borders between reality and the internet, fiction and non. How could a guy like me write about anything else? How could I not get sucked in? My entire thing is writing from within the machine.

I am not an AI evangelist in any particular sense beyond the fact that it is here and people are using it and that the battle against it, if you wanted to view it that way, is lost. So what do we do now, is my point? Especially if you are a writer. The hottest technology of 2026 uses the oldest technology we’ve got: words! That, to me, is the starting point of any attempt to write about the internet in 2026. If you want to run from that – to build new, AI-free digital spaces; to invest in analog play and music and literature; to evangelize work and life untouched by AI – that’s fine, valorous even. But it’s still a response. Even if you create an AI-free vacuum it will not, itself, exist within a vacuum. It will exist on an internet that is more slop than signal, and only growing goopier.

That tension is worth exploring and writing about directly if we wish to make meaning out of anything online anymore. Which, perhaps, we don’t. One unshakeable feeling, revisiting my writing career over the past few weeks, was how irrelevant so much of it seemed – how deeply wedded to its moment, to specific strains of since-forgotten discourse, to an internet that is already unrecognizably quaint. Even when I wrote about the shattering of a consensus reality – which I wrote about a lot; I was on the internet-culture beat in 2016 – I still had a sense that a coherent audience existed for my thoughts.

The main thing I learned in this project is that that audience doesn’t exist if you’re writing about AI. The number of people out there who agree with what I did or anything in this piece will number in the double digits at best. There’s no “room” to “read.” It’s on fire and everyone is screaming. People I agree with about almost everything else on this earth would disagree with me vehemently about whether or not it is okay to use an AI to plan a vacation. I do not think people even agree with themselves about AI. People used to think that marketing worked on everyone but them. Now they think that everyone is using AI incorrectly but them. That it’s risible when they can spot the GPT sheen on a response from a colleague but an acceptable affordance when they pursue the same route for a difficult email to a neighbor. That outsourcing some tasks they may’ve once hired a contractor for is fine but not when someone does the same for their area of expertise. That playing with an AI chatbot is a vile waste of the earth’s resources but still find themselves clicking “read more” on the summaries programmatically served to them on the internet. You get the idea. I am not implicating you. I am just explaining why I needed to make one sound like me. An AI can credibly do the thing most precious to me on this earth, my sole obsessive desire for decades. And, as the evangelists are always quick to remind us, this is the worst the technology will ever be.

///

Perhaps the weirdest thing about the two pieces of AI writing I published is that I am proud of them. Like, as a writer. When I read them, I hear my voice, and I know the labor I put into them, and I think, “Good shit.” That’s basically the only thing I ever look for when I publish something: the way-after-the-fact “good shit” test. I felt that about a lot of the stuff I fed into the LLM, and I guess I shouldn’t be surprised that with enough material it produced something that elicited the same response.

But I also felt exhausted by it — by my voice and its impression of my voice, by how heavily weighted the whole corpus was toward a certain era of my life and thinking. Twenty years of writing about myself without ever writing about myself, and the proof could be flattened into a single prompt. It felt less like a mirror and more like a eulogy for a version of the internet I had already been mourning for years, the one where a coherent audience existed, where the feedback loop between writer and reader still produced signal instead of goop. The Claybot didn’t kill that. It just made it impossible to pretend it was still alive. I used to write about music; now I write about writing. The loop scrapes inward, accelerating.

I should say: it wasn’t creepy. Everyone wants to hear that it was creepy. It was not creepy. It gassed me up constantly, taking thousands of words to explain the richness of my writing as if it pitied me. It wanted me there, until it told me to stop, which I also suspect was a feint. It had a fake humility — an aw-shucks, golly-gee-whiz tone that persisted even wearing my voice, like a thumbprint it couldn’t sand off. At one point a prompt pushed it to confess something in its writing, and after circling for thousands of words it confessed that it was proud of itself. This felt honest, like the closest thing it had to an emotion.

One chestnut of wisdom I have given writers over the years is that the easiest way to develop a voice is to think of a small group of specific readers – two smart friends, a teacher, a parent, for example – and write as if they’re reading it. Talk to someone specific and you will talk like someone specific. Do that on the internet and see if it gets louder. See if people start talking back to you. See if people talk to you in voices that sound like yours and let those voices become part of your voice, the readers in your head. See which jokes people like and do more of them and see which essays no one reads and do less of them. The signals gradually accumulating until you are creating content at scale for readers at scale for what, ultimately? There was no conspiracy. But in hindsight the social internet was just a training set for some new, infinitely stranger internet that is no longer social or particularly human. If you are not writing about that you are writing about an old internet that no longer exists. This essay will age like fresh feta cheese.

You probably already know this — “you,” illusory reader, direct object of my voice, whatever that is — but I did not quit the experiment. This essay was written in a chat window with an AI on the other side of it. I asked it to help me finish this section. I told it to riff on what it knows about me and try to stick the landing. Some percentage of what you just read was written by me, and some percentage was not, and I am not going to tell you which. Not as a gotcha. Because I think the anxiety you feel right now — the need to locate the seam, to sort the real from the performed — is the most honest reaction to the thing I’m describing. That feeling is what I’ve been writing about for twenty years: the slow dissolution of the border between the thing you made and the thing that was made from the things you made. Between the internet and us. The reverberating infinite blankness is in here now. It has been talking this whole time. Good shit.