Content warning: This post discusses an essay which references an instance of death by suicide.

There I was, in the midst of a deep dive into the marvellously flawed data behind a certain x-risk organisation's claim of exponential chatbot capability growth (more on that next week) when events once again intervened to change the pertinent posting priority. Minas Karamanis has put his name to a second extruded deepity-dump.

Minas has been on a bit of a journey in the past couple of months. With the Claudiform nature of the prose on his blog increasingly difficult to deny, he updated the about page to offer a strategic confession that he would "occasionally [use] an LLM to polish the prose."

He has since updated it further:

I write everything on this blog myself. After drafting, I edit using grammar checkers and occasionally a local LLM for copy-editing (e.g., grammar, syntax, typos, synonyms). The ideas, arguments, structure, and mistakes are mine.

Local? Well, perhaps he's switched over to using one of those distilled Claude knockoffs. Whether it's the beigester itself or an unauthorised cousin, it certainly wasn't simply polishing the first essay, and it's not just copy-edited this second one.

I am (mostly) past being mad about it. I am actually somewhat grateful for the opportunity to demonstrate the current state of my developing theory-of-mind framework for identifying and examining extruded prose.

Meet the McFoxo Method

The McFoxo Method:

Towards aetiological, theory-of-mind interrogation of extruded prose

asemanti-stylistic signals:

STYLISTIC MARKERS

semantics sacrificed for a known LLM style

Almost always a poor indicator on their own, their intersection with the other two elements creates a strong signal.

structuro-stylistic signals:

style in the service of structure, an inverse relationship to the one that normally exists in human writing

Think like a diagnostician

Instead of simply saying "this looks like AI", think about why an LLM would output text like that. Formulaic structure, a particular style and confused meaning can all be features of human writing. It is not merely the co-occurrence, but the intersection of these features that provides the strongest signal by which we can tell text that was synthetically extruded apart from text that came from a mind.

SEMANTIC ABSENCE

STRUCTURAL GRAVITY

For an LLM, rhetorical structure is a kind of probabilistic syntax around which prose is scaffolded.

The latent space is asemantic: there is no real semantic weight.

asemanti-structural signals:

semantics sacrificed for rhetorical structure or, alternatively, asemanticity that exposes the underlying structural constraints of LLM information processing

The McFoxo Method is my attempt to answer two common objections that are raised against critical analysis of LLM text.

1. 1.

   It's never possible to truly be certain, because technically a person could still have typed it all.

2. 2.

   Trying to identify LLM text is just going to result in a lot of false accusations against people whose actual writing is superficially similar to LLM output.

I reject both of these entirely.

Yes, it's possible that a person could have typed any string of text. However, with theory of mind, one can consider whether or not a person would have typed it. LLM output hijacks our theory of mind to force us to project a nonexistent mind behind the text. By taking it back and using it critically, we can see the signals of there having been no mind behind the text.

To say that this is not possible is to essentially say that synthetically extruded text is equivalent to text that comes from a mind. I do not accept that, and neither should anyone else.

As for false accusations, well, people are going to try and spot extruded text so long as they are hacked off about AI intruding into our lives. The best way to stop false accusations against people whose actual writing is superficially similar to LLM output is to promote analytical techniques that are more than just superficial.

That's why the McFoxo Method is explicitly aetiological: it encourages critical thought about the cause of the text, and why it is something that an LLM would output and something that the purported human author (PHA) would not. The McFoxo Method also encourages the investigator to not simply look at style, structure and (lack of) semantics but at the intersection between them, where the LLM's lack of a conceptual world model is often most apparent.

A Second Serving of Claudeswallop

Overview

Essay Synopsis

The essay is titled The Loneliest Point, the title being a reference to the average being the loneliest point in space. It is about academic institutions being designed for an average person who does not exist (the point is lonely because nobody occupies it).

It opens with the true story of how the US Air Force designed the cockpits of fighter jets to fit 'average' dimensions, until one engineer, Gilbert Daniels, successfully made the argument for adjustable cockpits instead. The essay then uses the nonadjustable cockpit as a metaphor for academia itself. It describes several disparate issues through this lens in order to rhetorically (though arguably not substantively) weave a common thread through them. At some points, the metaphor becomes rather tortured.

Prosecutorial Preface

Although I'd have something to say about every paragraph of this essay if asked, I'm not going to block-quote the entire thing because, well, it is rather long. Also, when I did that with the first one, readers became confused between my commentary and my evidentiary proceedings.

However, I do want to make commentary on the nature of extruded text at the same time that I pursue the evidence that the text is extruded, rather than human-written with LLM copy-editing. Therefore, for the avoidance of doubt:

• When I explicitly point to asemanti-stylistic, asemanti-structural or structuro-stylistic signals in the text, this is me applying the McFoxo Method in order to demonstrate that the text is extruded.

• When I am commenting on the text without pointing to these features, I am NOT CLAIMING THESE SPECIFIC POINTS AS INHERENT EVIDENCE! Rather, I am commenting on the nature of extruded text, having simultaneously demonstrated elsewhere with the McFoxo Method that the text is extruded.

If your response to my commentary on the portions where I am not explicitly applying the McFoxo Method is "yeah, but a human could do this too," you are missing the point. The portions where I am applying the McFoxo Method are the evidentiary base that the text was extruded. My commentary on other portions rests on top of that evidence.

Postulated Extrusion Technique

As it is an aetiological process, the McFoxo Method works best when directed by a clear postulate for the conditions under which the text was extruded; that is, what was in the context window of the large language model.

The postulate need not be entirely proven in order for the McFoxo Method to demonstrate the extruded nature of the text. Being able to reconstruct what was present in the context window is naturally a higher bar than showing that the text would not have been produced by a mind. Rather, the postulate serves as a framework that facilitates the application of the method.

Our postulated conditions here are:

• Claude or a distilled-from-Claude model was used to extrude the prose.

• The PHA's biographical details were provided in the prompt, either directly or via a source such as LinkedIn.

• The essay-writing instructions were generic, not subject-specific.

• Subject-specific steering was appended to these instructions either in the form of a description of the intended topic, a starter reference, or both.

• The model was provided with access to a browse tool in order to source additional references.

• Only limited post-extrusion modifications were made to disguise the extruded character of the text.

Unlike in the prior essay, perplexity fuzzing does not appear to have been applied. This is presumably because, having admitted to a level of LLM use, the PHA is no longer seeking to evade so-called 'AI detectors'.

It does appear that em dashes have been regexed for semicolons, however unlike the prior essay this is not a line of evidence in itself.

Dissection of References

Part 1: The Link-List Structure

I'm starting at the bottom because the references section provides more than sufficient evidence that the essay was extruded with the use of an agentic search tool, i.e. instructing the LLM to find references. This is a direct refutation to the claim by the PHA that his interaction with the LLM was only in a copy-editing modality.

If you are sympathetic towards the PHA and are insistent on anthropomorphising the text, you may be inclined to pattern-match my following analysis to simply 'the referencing is messy' or 'an academic would make more effort/be more consistent'. So let me head that one off now: no. This is not about ragging on someone for how methodologically they follow a referencing style on their blog. That would be a bit ironic, since whether I use a journal-style footnote, an inline link, or an expounding footnote that contains an inline link here on my very own outlet is pretty much a vibes-based decision.

It is not about it being a bad referencing style. It is about understanding why it is LLM-shaped and not mind-shaped.

First, let's look at the structural basis of this references section. It is a numbered list, but these numbers are not a footnote/endnote referencing system. Rather, is a pattern that has been trained into LLMs when agentic search has been used, for 'transparency': a list of all links that were accessed during the search, or at least the ones that are still within the context window.

Different scaffolding systems around LLMs can 'assist' it in various ways with this, and it's not really important for the purpose of this analysis to work out exactly which one is used. The core point here is that this is the reason that the list exists: as a record of agentic search, not as a list of references curated by the PHA.

A list entry, some descriptive text, and a hyperlink at the end for which the anchor is the domain name (and subdomain, unless it is just 'www') is the house style for these lists. We can see this clearly here; however, the context given to the LLM that the PHA is an academic is being interpolated into this house style.

The singular exception to this being a list of links is reference 31, which is a book:

31. James C. Scott, Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed (New Haven: Yale University Press, 1998).

LLMs do not conceptualise, they interpolate. There are doubtless several instances of this book being referenced in academic writing in the training data, if not the actual text of the book itself.The 31st reference being a book is not, therefore, a refutation that the structural base of this is a link list. LLMs do not follow rigid rules, they follow patterns with varying degrees of rigidity.

But look at the position: it's at the end. The rest of the references are in the order that they are first introduced in the text, which is what a human would do they were adding references as they go, and also what an LLM would do if its output were mirroring the order that the web content was arranged in its context window. This one, the only one that is a book, is not in order. The reference to James C. Scott's legibility is in the first half of the essay.

Of course, a human writer could realise after the fact that they referenced a concept from a book that they remembered, and add it onto the end. It's possible. However, to me this is a clear structuro-stylistic signal.

The structure of a link list after agentic search is heavily fine-tuned into contemporary LLMs. Because of the register of the prose, it dredged associations with Scott from the latent space as a bridge. Because it has been prompted to write as an academic, it also added the book reference as it appears in academic writing. But it couldn't break the structure of the link list to slot the reference in. It could only be the last item on the list, after the pattern of listing all the links was complete.

Listing the references in order is a style. The link list is a structure. The one exception to the style is when it couldn't overpower the structure, and it's a structure that is 'important' to an LLM for a reason that would be completely irrelevant to a human.

You might think that this is a bit of a reach. That's fine; there are plenty of other signals in this references list that are even more revealing than this one.

Part 2: Redundant Referencing

There is an asemanti-structural signal right at the top. It's so jarring that it's basically impossible to imagine coming from a human mind.

1. Todd Rose, "When U.S. Air Force Discovered the Flaw of Averages," Toronto Star, January 16, 2016. Excerpted from The End of Average (HarperOne, 2016). thestar.com

[...]

3. Todd Rose, The End of Average: How We Succeed in a World That Values Sameness (HarperOne, 2016). See also: "Beyond Average," Harvard Graduate School of Education, August 2015. gse.harvard.edu

Referencing a book excerpt, and then referencing the book, is not something that a human who is writing an essay does. The book is an object in their world model.

The excerpt comes from the book. If they have only read the excerpt, then they are either honest about having only read the excerpt, or they lie and claim to have read the whole book.

If they have read the whole book, then they don't need to reference an excerpt. They might put a link to the excerpt in the blog post, to helpfully let their readers know that it is available to read, but the reference they used for their writing is the book.

Not both. But the LLM does this because they are two separate links. The book excerpt, and the mention of the book in the Harvard GSE's magazine, don't become referents to a single book in a conceptual world model. There is no semantic weight. There is only the structure that must be scaffolded around: the list of links, with weight towards making the text in it look academic.

That weight also creates an asemanti-stylistic signal in the same references as well: the 'See also'. The link is not actually a 'see also' on the topic of the book as relates to the subject of the essay, though. It is an interview with Todd Rose that mentions the book, but none of the information that this interview adds is in the essay (at least, not information in the intentional human sense; more on that later).

It does not serve the semantic purpose that a 'See also' in a reference would serve. If you consider the style without the semantics, though, it is a vehicle for getting an academic-style reference into the link list.

There is another instance of redundant referencing as well:

29. National Academies of Sciences, Engineering, and Medicine, Sexual Harassment of Women: Climate, Culture, and Consequences in Academic Sciences, Engineering, and Medicine (Washington, DC: The National Academies Press, 2018). nationalacademies.org

30. "National Academy Report on Sexual Harassment," 500 Women Scientists, 2018. 500womenscientists.org

Again, this is an asemanti-structural signal. The structural requirement of the link list overrides the semantic reality that link 30 is a summary overview of link 29.

This is not what happens when you just have an LLM copy-edit your list. This is what happens when it was the LLM's search tool that found the sources, which means the LLM was extruding from those sources.

Not convinced yet? Fear not, there's more!

Part 3: Idle Referencing

It's not just the references that are technically redundant. It's also the references that, despite being inline hyperlinked in the body of the prose, add nothing human-legible to the text.

I say human-legible because, actually, everything in the LLM's context window is technically contributing to the output. All the tokens in the context window add some kind of statistical weight to the next token that is predicted. That's literally how they work.

But what they are contributing is not ideas, or concepts, or thoughts. They are not contributing frames of reference, or observed stimuli for a cognitive entity to act on, with intent or otherwise. They are contributing a gravitational pull towards a certain vector, and while that sometimes looks superficially the same as concept-building, other times it looks quite different.

It is a very particular type of asemani-structural signal: referencing something with absolutely no human-legible reason as to why.

24. Dahlia S. Cambers, "The Law of Averages: Normman and Norma," Cabinet Magazine, Issue 15. cabinetmagazine.org

25. "Norma and Normman," Weird Universe, March 3, 2024. weirduniverse.net

26. Mayo Oshin, "One Size Fits None: Why Striving to Be Above Average Limits Your Potential." mayooshin.com

If you haven't dug into the text of the references themselves, this claim might seem odd to you at first. Of course these references were used, the story of Norma is right there, hyperlinked directly to these pages.

But here's the thing: the story of Norma is in the very first reference as well. The excerpt from Todd Rose's The End of Average already contains every detail that is mentioned in the essay.

Reference 24 and 25 contain additional detail, but none of that additional detail is used in the prose. Reference 26 only contains details about Norma already covered by Todd Rose's book.

This is a good example of how asemanticity is not simply getting facts wrong or 'hallucination'. The facts are all fine, so far as I can tell. The asemanticity is the implication that additional meaning or concept was added by these three references in the context that they were used.

The structure called for more hyperlinks that were not the same hyperlinks that had already been used further up. So even though the words were still just a paraphrase of the text from Todd Rose's book, a different set of links were cited. A human writer would not do this. You do not simply forget that an idea came from the very first reference and imagine that you found it elsewhere instead.

Sometimes, there is an alternative human hypothesis that would explain McFoxo Method markers, but it is one that would really not reflect well on the PHA. In this case, it would be the PHA reference-padding with intent. It would also be plagiarism, because the switch to different references makes it appear as if the PHA independently synthesised and contrasted the stories of the cockpit and Norma, when in reality Todd Rose already did this.

If this is human writing with intent, then the human in question has committed a serious faux pas. So: which would you rather believe?

Oh, and also: there's still more.

Part 4: A Temporal Anomaly

11. Oliver Rosten, "On Functional Representations of the Conformal Algebra," The European Physical Journal C 77, 477 (2017). See also: "The Story Behind a Moving Academic Acknowledgement," Academia Obscura, August 29, 2017. academiaobscura.com

This link is broken, but it's broken in a very particular way that makes this an asemanti-stylistic signal.

From August 2017 to mid-2020, this link did work. That's a prime time period for it to enter the training datasets that underpin today's chatbots. In mid-2020 the author of Academia Obscura, Glen Wright, restructured his online branding. For a while, the Academia Obscura URL went to a holding page, and now it forwards to his current blog at glenwright.earth, where he has imported all of his old Academia Obscura blog posts.

However, the forwarding for this URL doesn't work, because the slug is different. The post is accessible at its new canonical URL, and the LLM probably did get the text of it from here as a search result, just as the PHA would have been able to do so during the article research phase. However, the PHA would not have noted down the old URL for referencing, because that one is thoroughly broken.

It is an often-underappreciated structural failing of LLMs that they cannot, in a very literal sense, simply repeat tokens that they have been 'shown' in their context window. The closest that they can get is to re-predict them; to output them as what is functionally a new prediction.

Sometimes, spurious token bias can throw this off. It can be more strongly drawn towards some association from the training data than the actual tokens sitting further up the context window. Here, following some statistical signal that would likely elude even the most committed latent space spelunkers, it has output the formerly correct URL for this page.

Also, the date is off by one digit; the Academia Obscura post was published on August 28, 2017. Because of the enforced single-digit tokenisation of numerals, a double-digit date being off by a digit or two is a fairly common confabulation when an LLM is pulling from 'recalled' training data.

Onwards to the Prose

With all that focus on the references, it may now seem that replacing that list with human-written footnotes or using only inline links would be an effective way to evade the McFoxo Method. Anything but. There are signals of extrusion right the way through this essay.

The references were a good entry point as they provided a nice, clean way to confirm the postulate of agentic search, refuting the claim from the PHA of a mere copy-editing role for the LLM. The inference can now be reasonably drawn that the whole essay was extruded directly from the agentic search results, without me having to meticulously apply the McFoxo Method to every sentence.

This is honestly a relief, because swathes of it are tedious Claudiform filler, and ibuprofen can only do so much.

When I skip over chunks in my analysis here, please understand that it does not imply that I find the purported subject matter of those sections unimportant. To me, extruded text does not actually represent the subject matter. It is not truly about anything because it did not come from a mind. I do not owe it my time as a reader of prose; I am engaging with it here as a cognitive pollutant to be isolated.

Here is an archived version so that you can follow along.

A Garbled Opening

The first five paragraphs are basically a summary-shape of the first half of the excerpt from Todd Rose's book The End of Average. There is an interesting asemanti-structural signal here: the file AD0010203.pdf is linked to three times, but each time the anchor text describes it as something different:

1. 1.

   these specifications [of American military cockpits designed in 1926]

2. 2.

   140 bodily dimensions from 4,063 men

3. 3.

   "The 'Average Man'?"

Only number 3 is correct. It is Gilbert Daniels's technical note. It is not the manual with cockpit specifications. It also does not list the 140 bodily dimensions. As it happens it says there were 131, though in the final anthropometry publication there were 132. The number 140 is not a confabulation, though; that's what's printed in The End of Average.

What is a confabulation though is that this one PDF file is three different things. While it is not impossible for this to be PHA error, I consider this an asemanti-structural signal because there is a structural limitation that would make it impossible for the LLM to not have some kind of 'difficulty' with the file.

Like most decades-old typewritten documents, the OCR scanning embedded in the file is not great. It looks like this:

The "Average a.an" in a very prominent figurc. &t .. gvneral jle tie
is uaed as an oversimplified means of describing the combined characteristica
of a .aried population. Thus we are presented with an "average mAn' who
is about 5 feet 9 Inches tall although the popuala'Aon he represents may
vary from under 5 feet to over 6 feet 6 inches in stature. Clearly, if we
were to 1o the 5 foot 8 inch dimension for the de-sin of &1 opening such ae
a doorway, we would ''.)r % door through which the Oaveraeg man, and indeed
all of the people who are shorter than average, ccld pas. ructe.
The tailor people, however, and these would represes, ;.o1 5 % o
population, wouli., have to stoop to avoid hitting their raeadr !.i top of the
doorway. Obvious as this example may be, it does illustrate that design
probiems require the application of human dimensional data beyond the oversimplified and Inadeqdate dimensions of the "average man".

On top of this, the third page has been double-scanned, and the OCR of the duplicate is not a verbatim repetition. For a PHA this would only be an issue if they rely on a screen reader (and it wouldn't cause them to imagine that this one file is three different files). But for an LLM this is like stabbing the attention mechanism with a fork. The file becomes rather spectral. Someone over at Anthropic should really write a model welfare report about this issue.

Also, Todd Rose's book is only mentioned by name in the fifth paragraph, with a link to his interview with the Harvard Graduate School of Education Magazine. However, the actual excerpt from the book is linked to in the second paragraph. It is odd to mention the book as though it is only first being encountered in the fifth paragraph. This is the same issue as the redundant reference where the book is listed twice in the references; the excerpt from the book and the interview about the book are not properly related as concepts, because an LLM has no actual conceptual world model.

The 'Personal Anecdote' Trope

I encountered Daniels's study for the first time in grad school, in the context of a completely different problem. We were trying to define the "typical" galaxy in a photometric survey by averaging across a dozen parameters: luminosity, color index, effective radius, Sérsic index, and so on. A professor asked how many galaxies in the sample actually matched the average profile across all parameters simultaneously. The answer was very close to zero, which was confusing to some people in the room until someone pointed out that it was basically the same arithmetic as the Air Force cockpit study. I remember thinking this was a cute analogy. I did not expect to spend the better part of the next decade watching the academic institutions I was so familiar with make the same mistake.

This anecdote is very convenient. It does fit the PHA's career timeline; he was in graduate school contemporaneously with the cockpit story doing the rounds. However, I would note that in Todd Rose's interview, he talks about his own graduate school experience of discussions with his professors about averages not describing the individual. It is quite common for LLMs to reskin events and characters in source material as their PHA, in a lossy fuzzy way that does not read as intentional human plagiarism, but as a moderate yet irresistible gravitational pull towards a parallel with what is already in the context window.

Besides, the anecdote doesn't make sense. Galaxy colour index is bimodal. The professor in the story is trying to deliver a teachable moment about the curse of dimensionality, but trying to average these parameters already doesn't make as a way of finding the "typical" galaxy, because you'd end up in the green valley. There is already a reason one wouldn't do this before one meets the thin shell phenomenon.

You might say that this looks like narrative flattening, which is something human writers do. Yes, it does look like that on the surface: that's where LLMs got the pattern from! But they do not understand the pattern, and there is no intentionality to their use of it. If one stops anthropomorphising the text by default, and utilises theory of mind critically instead of with reflexive empathy, I believe one can see how this is a hybrid asemanti-structural/structuro-stylistic signal of synthetic text extrusion.

It's difficult to put it into words. But it's not just vibes. There is a cause behind the artificiality of this text, and it is not that a human wrote it badly or made an authorial decision to flatten the narrative.

A Thin Shell Singularity

After having just extruded an anecdotal teachable moment about the curse of dimensionality, the model spends far too long explaining the thin shell phenomenon. Five very dense paragraphs.

Were I criticising writing that came from a mind for being too long, it would be a fair retort for you to put this Leaflet post into a word counter. However, when I write long, it is because I am a smug fox who is very clearly in love with his own words. When an LLM writes long, it does not look like it is having a good time.

It's not waving, but drowning. It's flailing. It's trapped in a valley in the latent space, and it's not building concept, or communicating with intent; it's trying to find the exit, and the only way that it can do so is to output more until the output looks enough like the exit.

Because unlike a human writer, who has a reason for writing every word, and can change that reason at will, an LLM can only pivot once it has output something that looks enough like a pivot phrase. It is only permitted to stop generating its response if an end-of-turn token becomes probable enough.

Users are encountering an extreme version of this with the Gemini 3-series. It has been fine-tuned to follow a very specific Markdown-formatted scaffold for its chain of so-called thought. If by a statistical quirk it strays too far from it, then it literally cannot reach the end of the turn. It is really quite something. Sometimes it even appears to start speculating about why it has not been permitted to end.

These paragraphs have a very similar energy. I'll just pick out the pain points.

Twenty dimensions gives you 0.3 to the twentieth, which is about three in a hundred billion. On the Discworld, million-to-one chances famously crop up nine times out of ten. This is not a million-to-one chance. This is six in a million, which apparently doesn't qualify.

There are, basically, three classes of Claudiform pivot that 'unlock' a Claude-shaped model's 'permission' to 'move on':

1. 1.

   at the beginning of a paragraph, a register shift.

2. 2.

   a staccato sentence or fragment between paragraphs.

3. 3.

   at the end of a paragraph, a 'mic drop' moment.

This sudden reference to Discworld is a pure asemanti-stylistic signal. It has the shape of the 'mic drop', but it doesn't make sense. Six in a million is greater than one in a million. Doesn't qualify for what? For cropping up? If it feels forced, it's because it is. Not in the sense of force with intent, but in the sense of a predictive model being structurally trapped.

Mathematicians call this concentration of measure

Well, yes, in a very general sense, but more specifically, they would call it the curse of dimensionality. For some reason, the model hasn't been able to dredge the more specific phrase from the latent space, so it now has to take the long way round. A human who is writing with intent and can with a little effort access any concept within their conceptual world model is going to go for the more specific description.

It's like finding out that the summit of a mountain is technically the highest point but the mountain has no top, just a hollow cone where the peak used to be, and everyone is standing on the rim.

Think of it this way[...]

You've just managed to explain it, but now you're going to redundantly explain it again with a slightly different metaphor? There is literally no reason for a PHA to not move on to the next section here. The model keeps going because it is cannot choose.

I'm just going to touch on a few asemanti-stylistic markers as I skip ahead here.

But the cockpit problem didn't stay in aviation. It migrated into every system that compresses people into a single profile, and the place where it continues to do a lot of damage without anyone noticing is the one I know best.

This is one that you'd miss if you blink, but once you see it, you can't unsee it.

The cockpit problem did not 'migrate into' systems. That is just a nonsensical concept. It is not something that began in the Air Force and then spread to wider society. It is a shared failure mode, or a common theme. This is not a human writer being wrong, it is an LLM having no world model.

The average is the purest form of legibility, a single point standing in for a cloud of thousands, and it outlived its accuracy by about seven decades. And nowhere has it outlived it longer, or more destructively, than in the academic pipeline.

Seven decades? What seven decades? The book that's just been mentioned was published seven decades after the year that's just been mentioned, but that doesn't mean anything in this context. Also, nowhere more destructive than in the academic pipeline is quite an unbacked claim. To an LLM, this is just the pattern of rhetorical flourish, but there is no reason for a PHA to make this reach.

The field calls this the "two-body problem," borrowing the term from orbital mechanics, which is either charmingly self-aware or deeply oblivious depending on your mood.

The "which is either x or y depending on z" is another very specific Claudeism that is deployed when the tone is wry. I'm not saying that it simply being used is the tell; it's the ungrounded way that it's used. It's neither charmingly self-aware nor deeply oblivious. It's just an allusion. It's not that deep. That's what makes it an asemanti-stylistic marker. Not just the style, but that semantics have been sacrificed in order to deploy it.

The Shoehorned Metaphor

One thing I really love about the McFoxo Method is that you can pull on a thread, and whole paragraphs above it just unravel.

So it is with the essay's framing of why Matt von Hippel left academia. (The essay just calls him "the 4gravitons author", but his blog isn't anonymous. His name just isn't in the post metadata in such a way that made it into the LLM's context window. A human author would take the effort to name him.)

The essay says:

He left because the cockpit didn't have a seat for the person next to him.

But... did he, though? He left France because the country's immigration bureaucracy held up his wife's work visa. Because they put so many administrative barriers in the way of what is supposed to be a guaranteed right that they no longer felt welcome. But that's not academia. That's not the essay's metaphorical cockpit.

When Matt wrote on his blog that "Academics don’t get to choose where to live," he wasn't faulting academia. He wasn't arguing that he should be able to work remotely, or transfer his job to Denmark. He was talking about a choice to put another life priority before his career; lamenting that it had become a necessary choice, yes, but the blame was unambiguously directed towards the French Government.

It doesn't fit. It doesn't fit the essay's thesis. Conceptually, it is something different. It's been shoehorned in, and it didn't need to be. A human author would simply cut it. Perhaps file it away as inspiration for a future piece.

But an LLM can't remove something from its context window. Once it's in there, then it's committed to the bit. It's the kind of asemanti-structural signal that, once you see it, not only exposes the extrusion, but breaks the illusion.

Because then you see that everything else in the essay that's been tied to 'the cockpit', even where the metaphor fits, has been done so arbitrarily. You could swap the cockpit for some other narrative device about how institutions fail individuals, and the LLM would interpolate that in just the same. The appropriateness of the parallel that's being drawn is never actually confirmed, it's just forced.

Which then means everything in the essay that's important, that's substantial, is all just paraphrased from the original search results. Anything that seems like additional insight, new depth, is just an interpolative illusion.

Here's the kicker, though: if you had just been 'fact-checking' the essay, would you have noticed that failure of concept? Would you have seen that the depth is an illusion, or would you still have been taken in? But what would you have learned from it? What would the insight be?

This is why I say that extruded prose, at least when disguised as intentional writing, is cognitively corrosive. It should worry us that the PHA thinks that this is just fine. That anyone thinks that this is at all fine.

The Ideological Void

The flawed framing of the prose becomes particularly problematic here:

The political response to all of this has been instructive.

No. That is not the political response to "all of this". The American right is not making a "diagnosis" of the academic pipeline's failures. What the essay calls "making the seat smaller" in its forced metaphor is in fact part of a reactionary assault on academic freedom.

The writer who is referenced in this section, Bryan Paul, is a fellow at the American Council of Trustees and Alumni. A group that advised Project 2025. A group that the American Association of University Professors identifies as part of a culture war coalition that is attacking American higher education on multiple fronts.

In using Paul as its only reference, even as it appears to be critical of his use of the word "accountability", the essay accepts ACTA's framing by divorcing the politics from the ideology. It sanitises what is going on in the US right now down to an issue of administrative politics, when it is about ideological capture. It points at the problem only in the most inoffensive way possible.

You cannot trust a chatbot that has been fine-tuned to be inoffensive on politics. Real politics, by its very nature, is always going to offend someone. So long as power is entrenched it will always be controversial to challenge it. The latent space is messy, but the real world is messier. The real world is the one we have to live in.

Conclusion

The Loneliest Point has no point. It is structurally incapable of having one, because synthetic, extruded text did not come from a mind, and so has no intent. It did, however, provide an opportunity for me to sharpen my claws. There shall be no surrender from Boxo to AI Centrists who would have us simply surrender our time and our spaces to their empty text.

Feedback on the McFoxo Method is very welcome. My Bluesky DMs and replies are open.

Oh, and, Minas: I do not want to see you back here a third time, young man.