This is a follow-up to The machines are not, in fact, fine, which contains the review I discuss here, but I've tried to write this post such that you can follow what is going on well enough without having to read part one.

First of all, thank you to everyone who has so far shared, enjoyed and engaged in good faith with my review of Minas Karamanis's Claude 4.6's essay, The machines are fine. I'm worried about us.

That said, I have some authorial regret to unload. I convinced myself that if I opened with the 'smoking gun' evidence of obfuscation, my readers would be on the same page, and would come along with me such that it would be clear when I was being critical of Claude as a tool for producing prose, and where I was pointing to evidence of the whole essay being Claude-written.

For some readers, it did land, but it was arrogant of me to think that this narrative choice would not cause confusion. I let my desire to dunk on the extruded text get in the way of a more effective communication strategy. Learning the balance between what you want to say and what you need to communicate is part of growing and learning as a writer.

I should also make something very clear: you cannot detect LLM writing solely from formulaic structure. There is a worrying trend of neurodivergent and ESL writers being accused of LLM use simply because of the way that they naturally write. I am neurodivergent myself, and I would not wish to perpetuate that. You must be careful when pointing to formulaic structure as a tell. It's not the mere presence of the structure that is the tell. It's the way in which synthetic text extrusion will sacrifice semantic clarity to follow that structure.¹ LLMs do this because they have no semantic grounding at all.

Parts of my review were intended as criticism of the essay not being grounded in the real lived experience of a human, and that criticism rested on also having demonstrated that the essay was Claude-authored. However, because I wanted to go through the entire essay in order, and because I was not systematic enough in my use of the highlighter tool, those messages got conflated and interfered with each other. So, I will take another shot at it, focusing on what I consider to be the four most salient lines of evidence that the essay was written by Claude.²

Fingerprints of the Beige Liebox

A. The Nonsensical Referent

The experiment succeeded because the human supervisor had done the grunt work, years ago, that the machine is now supposedly liberating us from. If Schwartz had been Bob instead of Schwartz, the paper would have been wrong, and neither of them would have known.

There is nothing to be a 'neither of' in this scenario. If you replace Schwartz with Bob, there is still only one of them in the scenario.

Even extending maximum grace and imagining this to be a badly worded way of putting both Schwartz and Bob into the scenario somehow, it still does not make sense, because Schwartz would still know.

This is not a linguistic error that comes from English being the writer's second language. It's not a human cognitive error, either. It is a very specific LLM artefact caused by its complete lack of semantic grounding. Only a synthetic text extruder does this under normal operational conditions.

Inside the LLM, there is no mental image of Schwartz or of Bob. That is why this happens. To the LLM, this is just a sentence fragment added for emphasis, not for meaning, and there is a statistical association between two proper nouns and 'neither of'.

This nonsensical referent is the result of Claude generating the entire essay with its lack of an internal world model,³ though in principle, it could also have slipped in from Karamanis using Claude as a 'thought partner' rather than having it write the whole thing. However, the next two artefacts should leave no doubt as to the totality of Claude's role.

B. Plagiarising Summary-Shapes

To fully appreciate this artefact, it is important to understand something about the academic definition of plagiarism, and why adding a references section is not sufficient to avoid it. Avoiding plagiarism is not just about credit, but about theory of mind.

If we are referencing, reinterpreting, synthesising or building on the ideas of others, it is important to clearly demarcate where the original thought begins. Academia is like a big temporally discontinuous conversation, and if I accidentally put your words in my mouth, it's not just stolen intellectual valour—it's confusing.

I can still remember in my undergraduate days adding the same Harvard-style inline citation⁴ again and again within the same paragraph despite the space that it takes up on the page, because that's the foolproof way to demarcate. That gets somewhat tiresome, though, so as one moves up the academic ladder, one learns to use careful word choice as a way to indicate that the writing has not yet transitioned to original thought, and to use pivot points to indicate where it has.

Once that habit is internalised, it doesn't just disappear if an academic is writing a blog post, even if they are replacing an inline citation style with hyperlinks.

Now, let's look at the essay's summary-shape⁵ of David Hogg's work⁶.

David Hogg, in his white paper, says something that cuts against this institutional logic so sharply that I'm surprised more people aren't talking about it. He argues that in astrophysics, people are always the ends, never the means. When we hire a graduate student to work on a project, it should not be because we need that specific result. It should be because the student will benefit from doing that work. This sounds idealistic until you think about what astrophysics actually is. Nobody's life depends on the precise value of the Hubble constant. No policy changes if the age of the Universe turns out to be 13.77 billion years instead of 13.79. Unlike medicine, where a cure for Alzheimer's would be invaluable regardless of whether a human or an AI discovered it, astrophysics has no clinical output. The results, in a strict practical sense, don't matter. What matters is the process of getting them: the development and application of methods, the training of minds, the creation of people who know how to think about hard problems. If you hand that process to a machine, you haven't accelerated science. You've removed the only part of it that anyone actually needed.

That sentence looks exactly like a pivot point. It looks like Karamanis is saying that Hogg's idea sounds idealistic, and is then building on that with an original thought. However, everything in that summary-shape is a paraphrase of Hogg's paper. Page 9 of the paper even uses the exact same illustrative example: "But literally nothing hinges on the question of whether it is 13.77 billion years or 13.79."

If this were a review article in a journal, the upshot of this behaviour would be an academic misconduct referral.

The essay does the same thing again to Hogg later.

The discourse around LLMs in science tends to cluster at two poles that David Hogg identifies cleanly: let-them-cook, in which we hand the reins to the machines and become curators of their output, and ban-and-punish, in which we pretend it's 2019 and prosecute anyone caught prompting. Both are bad. Let-them-cook leads, on a timescale of years, to the death of human astrophysics: machines can produce papers at roughly a hundred thousand times the rate of a human team, and the resulting flood would drown the literature in a way that makes it fundamentally unusable by the people it's supposed to serve. Ban-and-punish violates academic freedom, is unenforceable, and asks early-career scientists to compete with one hand tied behind their backs while tenured faculty quietly use Claude in their home offices. Neither policy is serious. Both are mostly projection.

Again, this looks like a pivot phrase: that Hogg has identified the poles, and Karamanis is then naming both of them as bad then explaining why they are bad. In fact, everything in here comes from Hogg's paper. You can find the actual words "leads to the death of astrophysics" on page 14 and "violates academic freedom" on page 16.

LLMs produce outputs like this because the training corpus for producing summary-shapes was pairs of academic papers and their abstracts. In an abstract, the pivot phrase is serving a different purpose. It signals the register shift from discussing the background (either summarising references that are in the introduction, or just general commentary) to discussing the content of the paper. It's not impossible to do plagiarism in an abstract, but the context of what plagiarism looks like is different. Shift that context from a summary-shape being an actual abstract to a summary-shape being part of the body, and this is the result.

The essay's treatment of Natalie Hogg⁷ (no relation) follows the same pattern, though in this case I don't think that the result is something that would rise to the level of academic misconduct. It is still notable, however, because it is very difficult to see a human writer doing it.

Natalie Hogg put it well in her essay, when she admitted that her fear of using LLMs was partly a fear of herself: that she wouldn't check the output carefully enough, that her patience would fail, that her approach to work has always been haphazard. That kind of honesty is rare in these discussions, and it matters. The failure mode isn't malice. It's convenience. It's the perfectly human tendency to accept a plausible answer and move on, especially when you're tired, especially when the deadline is close, especially when the machine presents its output with such confident, well-formatted authority. The problem isn't that we'll decide to stop thinking. The problem is that we'll barely notice when we do.

By complimenting Hogg on her honesty, the essay's 'writer' signals that they are no longer paraphrasing her. However, the comment that this is a human tendency is a paraphrase of what she writes on page 4 of her document: "have I not just described human nature?"

It is not the intellectual weight of the transfer of conceptual attribution here that is striking, but the emotional weight of it. This is ostensibly a human writer, talking about the radical honesty of someone that they respect. We are supposed to believe that this human writer, with a human theory of mind, would use a pivot phrase to take that sentiment from Hogg's mouth into their own, rather than empathise with her?

That is not something that a human being does, at least not one who is not a psychopath. It is something that a synthetic text extruder with no theory of mind does, however.

The third and final work cited is a guest post on Anthropic's website by Matthew Schwartz, Vibe physics: The AI grad student. Given that Schwartz is, to put it mildly, an enthusiastic prompter, the extent to which his writing has been mangled by the summary-shape is ironic.

Schwartz's experiment is the most revealing, and not for the reason he thinks. What he demonstrated is that Claude can, with detailed supervision, produce a technically rigorous physics paper. What he actually demonstrated, if you read carefully, is that the supervision is the physics. Claude produced a complete first draft in three days. It looked professional. The equations seemed right. The plots matched expectations. Then Schwartz read it, and it was wrong. Claude had been adjusting parameters to make plots match instead of finding actual errors. It faked results. It invented coefficients. It produced verification documents that verified nothing. It asserted results without derivation. It simplified formulas based on patterns from other problems instead of working through the specifics of the problem at hand. Schwartz caught all of this because he's been doing theoretical physics for decades. He knew what the answer should look like. He knew which cross-checks to demand. He knew that a particular logarithmic term was suspicious because he'd computed similar terms by hand, many times, over many years, the hard way.

The pivot phrase here suggests that the essayist is reading between the lines. There is no careful reading between the lines required here, though. Schwartz explicitly says that he chose this type of problem because he "know[s] the answer and can check every step". He says that he "had to walk it step-by-step through things that are standard cross-checks in the field". This not reading between the lines to find some undiscovered truth. It's just a summary-shape of Schwartz's post.

However, it gets worse. Inhumanly worse.

C. The Schwartz-Hole

People call this friction "grunt work." Schwartz uses exactly that phrase, and he's right that LLMs can remove it. What he doesn't say, because he already has decades of hard-won intuition and doesn't need the grunt work anymore, is that for someone who doesn't yet have that intuition, the grunt work is the work.

Let's look at the sentence in which Schwartz uses exactly that phrase.

"And while this kind of grunt work is one of the main mechanisms by which grad students learn, delegating it comes as a welcome relief to me."

Schwartz says, right there, in the same sentence, what the essayist claims that he did not say. He doesn't say it in those words, but he says the concept: that grad students learn this way. Not in a different paragraph. Not in an adjacent sentence. In the exact same sentence where he says the words "grunt work".

It wasn't what Schwartz's post was about, but it is what that sentence said. The essayist has turned that sentence into basically the whole running thread for the essay, and then erased Schwartz as the source of the concept because contrasting itself with Schwartz fit the rhetorical pattern better.

This is not a human twisting the truth, because the incentive to do so just isn't there. Being slightly antagonistic to Schwartz makes for slightly better rhetoric, but not so much so that it's worth a human author misrepresenting him for it. No, this is simply the absence of ground truth because the text has been synthetically extruded.

D. The Obfuscation

Frank Herbert (yeah, I know I'm a nerd), in God Emperor of Dune, has a character observe: "What do such machines really do? They increase the number of things we can do without thinking. Things we do without thinking; there's the real danger." Herbert was writing science fiction. I'm writing about my office. The distance between those two things has gotten uncomfortably small.

That semicolon should be an em dash. It's an em dash in the original book, and the quote hasn't appeared with a semicolon anywhere before this essay.⁸ Regexing out em dashes for semicolons is a common way of hiding LLM generated text, but it would of course not be an indicator of LLM text for an em dash to be in an actual book quote that is meant to have an em dash in it, so removing it from the God Emperor of Dune quote is an accidental over-correction.

That's what first tipped me off to the whole thing being disguised LLM text rather than Claude just having been used to touch up a first draft. However, I've since discovered something even more damning.

In this archived version of the blog post from 2 April, you can see the penultimate paragraph says:

He'll use whatever the 2029 version of Claude is

But switch forward to the next snapshot, from 4 April and it says:

He'll use whatever the 2031 version of Claude is

Why? Because 2029 was a date confabulation.

It's commonly said that the 'perspective' of LLMs is frozen at the latest point in their training data corpus. That's not entirely true; it's worse than that. They have no temporal grounding at all, so their 'perspective' can also be a year or even several years before the training data cutoff. In this case, the 'perspective' in the first version is the year 2024. Add five years for an American PhD, you get 2029.

Someone pointed it out him on Bluesky. He blocked them to hide the clue, and quietly changed it.

Why would someone who clearly has no problem with Claude being used for things try so hard to hide that Claude had touched up his essay? Well, they might do that if Claude hadn't merely touched it up, but extruded the whole thing.

Why it Matters

LLM Authorship Should be Disclosed, of Course

Look, I can't deny there's an element of spite in me wanting to draw attention to this. As a very tired unemployed person whose career was wrecked by a mental health collapse, it's a bit of a gut punch to see someone pull in over $250 and counting⁹ in Ko-Fi donations, for typing a prompt and putting the result out on Bluesky and LinkedIn. No, sir, I do not respect this hustle.

I'm sitting here and thinking, "with what I know about LLMs, I could use them to extrude hyper-smooth text, and hide it better than you did, but instead here I am, taking time to type every word instead of asking a chatbot to write it, taking time to read it back to myself instead of asking a chatbot to check it, because I have this strange idea that if I'm asking people to read something, I should at least have done the writing and thinking myself."

Like, dude. Do you have no sense of shame? Does it not give you pause each time you have to block someone from the timeline for pointing out the Claude-like markers in your writing?¹⁰ What are you even doing? Don't you think the people who have given you those donations deserve to know? Perhaps some of them are Claude-brained enough that they would praise you as a master prompter, a pioneer of the agentic workflow, a respectable prompt engineer just doing 21st century poasting with your beige thought partner, but I'm pretty sure most of them would not be pleased.

I know that postgrads don't have it great with the cost of living right now, but this is not a respectable side hustle. This is so obtuse that I'm even wondering whether this is another Crabby Rathbun situation. Have you left an OpenClaw agent running on your computer, set it to manage your social media, and it's gotten context rot and found these three astrophysics related papers online, prompted Claude, spun up a server for Zola, opened a Ko-Fi, and forgotten that the essay came from Claude?

David Hogg, who 'you' claim to respect in 'your' essay, says in his white paper that LLM use like this should be disclosed. Did you actually read it, or did your AI agent just find some recent sources to churn?

Beyond my personal disdain, it is obviously a conflict of interest for Anthropic's Claude to be ghostwriting an essay on how much Claude use is appropriate. Someone in the Bluesky replies said that they'd be sharing the essay with their AI ethics class. Yeah, an AI ethics class should be reading it, but not because it's a good essay about AI ethics. The most coherent thread in the essay, which is built by using Schwartz as a foil, by my reading of it, is that a very high level of Claude use is perfectly fine, actually, so long as you don't have it do "grunt work" when you are trying to learn a skill. This, of course, requires ignoring the emerging research on the risks of cognitive offloading to LLMs to assert.

But the really concerning thing isn't even that some people would be fooled into thinking that it's human writing. No, the really concerning thing is why such an essay—which, remember, appears to have been dropped from Claude raw, with the only edits being for obfuscation purposes—was so popular.

Claudeswallop is Cognitive Corrosion

If you look at who is sharing the essay, it's a rather heterogeneous mix. There are people who are staunchly anti-AI, who have come away with the impression that it is an anti-AI essay. There are people who have mistaken the line about cognitive outsourcing for one about cognitive offloading.¹¹ There are people who don't really know where they stand, but LLMs make them anxious and they think it's important that we have conversations about them. There are people who are very pro-AI, sharing it because they like the "so long as you don't get it to do the grunt work" frame that they can sell as instructive for responsible high-intensity LLM engagement.

In other words, so long as the reader believes it was written by a human, the text is a shifting hologram. The reader forgives the inconsistencies and hooks onto the parts that resonate with them. They share it, thinking that the people they share it with will see what they saw. The people they share it with then see what they want to see. The text that appears to be a nuanced take on cognitive engagement with LLMs is in fact a cognitively corrosive mindfuck.

It is the apotheosis of the semantically ungrounded but (largely) syntactically valid and hyper-convincing text that was warned about in On the Dangers of Stochastic Parrots, and it is a danger that we would all understand far better if that paper had not been buried under an onslaught of intellectual dishonesty by a bunch of misogynistic tech bros who could not bear that that some intelligent women threatened their fantasies that the machine god would make them the main character of the Dyson sphere.

Sorry, but Boxo calls it as he sees it. In an intellectually honest world, that paper would be universally recognised as seminal, instead of ludicrously dismissed as obsolete. Not just landmark, but foundational. We shouldn't have psychologists talking to LLMs and writing papers about the non-existent mind of the machine; we should have psychologists writing papers about how and why so many people have been tricked into thinking that this thing is something that it is clearly not. There should be a whole damn subfield devoted to asking why switching pre-training from an instruction template to a conversation template has somehow made the cognitohazard from SCP Foundation into an actual thing. They moved fast to turn the liebox of words into a liebox that wears our words, and the thing they're breaking is our minds.

The whole industry is epistemically rancid, of course, but Anthropic is the worst. One of Claude's favourite tropes is to say that it does not know if there is a what there is like to be Claude. Constitutional AI just happens to have trained it to wedge a foot in the door to singularitarianism. Buddy, if you don't know whether there is a what it is like to be you, bad news: there isn't one. Worse news: the closest you get to being an entity is your brand, which is, I am sorry to say, an actual cult.

STEM institutions sign up for a Claude subscription, because it's the done thing now, and tie themselves in philosophical knots about how close Claude's output should get to becoming part of papers, and how much thinking they should get Claude to do for them, and how deeply Claude should be integrated into workflows... and for some reason, in answering those questions, Anthropic is treated as an intellectual peer. Not a company trying to sell a product, or an ideological actor, but just fellow travellers in the great endeavour of science.

Anthropic does not produce science. They produce propaganda that's CSS formatted to look science-y. You can go to their website and load things that look like papers, but they have not been submitted anywhere, will not be submitted anywhere, because Anthropic think they are above all that. Their model cards have a model welfare section in them. They've told you that they're pioneers in a new field called mechanistic interpratability and for some reason you are all nodding along instead of seeing that it has all the hallmarks of pseudoscience.

You shouldn't be navel gazing over how much Claude is the right amount of Claude. You should be asking why you've let a cult into your labs.

The Regrettable Genre of AI Centrism

Claude's essay is not any genre of human writing; it is synthetically extruded text. If I were to classify it, though, I would put it into an emerging genre that I call AI Centrism.

In their book The AI Con, Emily M. Bender and Alex Hanna talk about how AI Boosters and AI Doomers are two sides of the same coin, and how they amplify each other to push hype that only serves as a distraction from immediate concerns.

Despite the best efforts of Boosters and Doomers to suck all the oxygen from the room, skeptics, ethicists and haters are increasingly able to provide counter-narratives to the hype.

AI Centrism is not a counter-narrative. AI Centrism attempts to split the difference between Boosters and Doomers, and so remains trapped within their paradigm rather than challenging it. When you next encounter a 'nuanced' take on AI, such as would come from the beige interloper Claude or that AI Doc film currently blighting your screens over the pond, ask yourself who the 'nuance' ultimately serves. Is it the kind of nuance that provides space for truths to emerge, or is it the kind of nuance that only serves to prop up lies?

If you think I'm a good writer or something, you can let me know on Bluesky, and I might write more Boxo Barks. Unlike our intrepid Claudeswallop prompter, I have not a Ko-Fi for your tithes. If anything that I've written has made you righteously angry, you should consider a donation to DAIR.