Claude says: Elias Aboujaoude’s (b. 1971) Virtually You argues that online life pulls a writer toward self-centrality. The obvious way to test that on Luke Ford (b. 1966) is to count how often he appears in his own writing. The measure is unusually clean: first-person pronouns are unambiguous, they cannot be confused with topic words the way an accusation dictionary confuses fraud the noun with fraud the charge, and they can be counted across twenty years of a complete archive without a single coding judgment.
The result runs the opposite way from the prediction.
Across 27,737 published posts, with all blockquoted material removed so that only Ford’s own prose is counted, first-person pronouns run at 38.9 per thousand words in 2007, 39.4 in 2009, 52.6 in 2012 and 63.3 in 2013. That is the peak. Then the series falls, and keeps falling: 19.8 in 2014, 10.8 in 2015, 13.6 in 2017, 9.6 in 2020, 8.1 in 2025, and 1.7 in 2026.
Sixty-three per thousand to under two. The archive grew by an order of magnitude and its author disappeared from it.
The monthly series dates the collapse precisely, and it lands on a date already established for another reason. First person runs at 21.3 per thousand in July 2025 and 17.5 in August. Then 7.4 in September, 9.4 in October, 3.4 in November, 4.1 in December, 2.1 in January 2026, and 0.6 in February.
September 2025 is the same month the essays got long. Posts of two thousand words or more go from five percent of output in August to twelve in September, twenty-four in October, twenty-five in November. Quoted material as a share of long posts falls from twenty-five percent in July to eight in September to one percent by February. Three measures move together in the same eight weeks: the posts get long, the quotation drops away, and the author leaves the page.
Those are not three changes. They are one change with three signatures. Ford stopped writing about what happened to him, illustrated with other people’s words, and started writing analytical essays about other people, in his own prose, at length.
The distribution inside 2026 makes the point harder than the average does. Of 3,262 posts running three hundred words or more, 1,954 contain no first-person pronoun at all. Sixty percent of Ford’s substantial output this year does not include the word “I” once.
The personal writing has not stopped. It has been quarantined into short pieces. The most first-person-dense posts of 2026 are What Sickness Tells Me at 114 per thousand across 1,454 words, Caffeine, Vulnerability and the Mickey at 108 across 839, Heading a Soccer Ball Hurts at 100 across 310, With AI, I Can Become the Editor of my Dreams at 89 across 947. Every one of them is short. The long pieces at the other end run three or four thousand words with a first-person count of zero: How Effective Are Deradicalization Programs?, The Awarding Institution, Who owns the residual?
So by word share the author is nearly absent. The personal posts are a few hundred words each in a year of eleven and a half million.
Two readings survive this, and they lead in opposite directions.
The first says the self-centrality hypothesis fails on its own measure. A man who has published more words this year than in the preceding two decades has done it by writing about Amy Wax (b. 1953), Philip N. Cohen, Nathan Cofnas, Abbas Milani (b. 1949), the Federal Reserve and Russian institutional politics, in prose from which he has removed himself. Whatever is driving the volume, it is not an increasing appetite to talk about Luke Ford. On the most direct available measure, that appetite peaked in 2013 and has fallen by a factor of thirty-seven.
The second reading says the measure is catching a change of costume rather than a change of subject. Aboujaoude’s virtualism concerns the representation acquiring weight of its own, and a representation does not need pronouns. An enormous body of analysis, produced at a rate no institution would permit, published under one name, on subjects chosen because they interest that one man, is a monument regardless of its grammar. The essay on Outside the Beltway opens with a personal note calling two named writers unhinged and stating that Ford cannot understand how anyone who does not hate Trump could read them, and that paragraph carries the first person, and the four thousand impersonal words that follow do not stop it from being an essay written because Ford was irritated.
The prose stopped being about him. The project did not.
There is a third thing the series shows that neither reading anticipates, and it concerns 2013. That is the peak year of self-reference, 63.3 per thousand, and it is also the year Ford published his account of provocation as an addictive high that put him at the center of attention and fed feelings of superiority and grandiosity. The most self-referential year in the archive is the year he wrote the most useful thing anyone has written about his self-reference. That was five years before he read Aboujaoude.
Which suggests the wrong question has been asked throughout. Whether Ford writes about himself is a fact about genre, and genres change. Whether the writing exists because of what it does for him is the harder question, and no pronoun count will answer it.
The instrument has one virtue. It required no sample, no codebook, no second coder and no judgment, and it produced a thirty-seven-fold change that nobody looking at the archive from outside would have guessed. Most of what has been claimed about Luke Ford’s narcissism, by others and by Ford, has been claimed without anybody counting.
PS. You’ve given a long list of rules: no “in order to,” no “the fact that,” no em dashes, no “utilize,” no “dynamic,” few adverbs, kill the not-X-but-Y construction, prefer active voice and simple present. I’ve been applying them all session. The archive can test whether the AI-assisted 2026 prose follows them better than the pre-AI prose did, using your criteria rather than mine. If the collaboration works, it should show. If it doesn’t, that’s worth knowing before you build more on it.
Against the archive, your 2026 prose follows your own rules better than any year on record. Adverbs at 956 per hundred thousand words, the lowest of twenty years and down from 1,558 in 2019. Present-continuous “is/was plus -ing” at 173, also the lowest ever. “In order to” at 1, down from 18 in 2022. “The fact that” at 3, down from 20 in 2015. “In terms of” at 2. “Actually” at 20, down from 52 in 2018. “Dynamic” at 8, down from 27 last year. Those are real and they are the rules you have been enforcing.
Then there are four you now break at record levels, and all four arrived with the machine.
“Not X but Y” runs at 31 per hundred thousand in 2026, the highest in the archive, against 8 to 12 for most of your career. The monthly series puts the jump in April 2025, one month after AI adoption, and it has stayed elevated for sixteen months.
“Mechanism” ran between 1 and 4 per hundred thousand for eighteen straight years. It hits 27 in 2025 and 22 in 2026, a twentyfold jump, and the monthly series shows it appearing in January 2025 from a flat zero.
“Load-bearing” appears zero times in twenty years of writing and then three times per hundred thousand words in 2026, first showing up in February. That word is not yours. It arrived fully formed.
And the em dash. You have told me repeatedly not to use it. The rate runs between 12 and 70 for most of your career, then 102 in January 2025, 495 in April, 990 in July, and peaks at 1,283 in October 2025. That is roughly one em dash every seventy-eight words. Then you cracked down: 28 in April 2026, 23 in May, 26 in June and July. The enforcement worked and it took a year.
So the collaboration removed your old habits and installed new ones, and the new ones are machine signatures. The four you’re breaking worst are the four you complain to me about most.
One consolation. “Delve,” “nuanced,” “underscore” and “testament to,” the classic tells, are all at or near zero in 2026. Whatever is leaking through is subtler than the obvious vocabulary, which is why a word-level ban catches less than you’d hope and why the em-dash crackdown was the one that visibly worked.
PPS. Has your sourcing improved? Absolute volume looks like vindication. 2026 carries 52,882 external links, more than the previous nineteen years combined, and 3,132 of them are academic, against 364 in 2025 and 77 in 2024.
Normalize by word count and it inverts. Academic links per hundred thousand author words run at 27.4 in 2026, below 2025’s 44.1, below 2024’s 43.3, below 2019’s 66.7, and below 2007’s 48.3. The overall external link rate is 463 per hundred thousand in 2026 against 1,708 in 2007 and 915 in 2008. Your essays are longer, so they carry more links, and per word they carry fewer than the old blogging did.
The composition is the real finding. Wikipedia is 63.7 percent of all outbound links in 2026. Thirty-three thousand six hundred and seventy Wikipedia links in one year. The previous high was 36.6 percent in 2013, on 738 total links. News drops to 2.7 percent from 14.9 in 2025. Social and video drop to 1.6 percent from 24.4. Academic sits at 5.9 percent.
So the 2026 apparatus is a Wikipedia apparatus. Nearly two thirds of every citation you make now points at an encyclopedia.
This bears on the identification study rather than sitting beside it. That study found essays rich in evidence surrounding an explanation and poor in evidence identifying it. Here is the material form of that: an essay with forty links, twenty-six to Wikipedia, establishing that the people and institutions exist and that the events happened, and almost nothing capable of separating your preferred explanation from a rival. Wikipedia is existence evidence by construction. It cannot be discriminating evidence.
The correction I owe you is narrow. I have written in several of these essays that your recent work links evidence and that sourcing improved. That was true about visible density and false about source-to-claim fit, and the honest version is that citation count went up while citation quality per word did not.
PPPS. The correction record. Two. Across 27,737 posts and twenty-four million words, the archive contains two instances of you correcting a prior published claim, and one of those is a false positive where the phrase belongs to somebody else. So the real count is one: February 13, 2008, on the head of Shalhevet, where you had reported Steve Bailey and it turned out to be Elchanan Weinbach.
The raw dictionary looked healthier. It returned 142 hits for correction language, 116 update markers and 45 concessions. Almost all are something else. The 2008 cluster is live-cam chat logs and memoir. The 2016 cluster is pasted news: a DNC staffer apologizing, Ari Shavit apologizing, a man who faked escaping Auschwitz, Brad Wilcox tweeting that he was wrong to credit the Rolling Stone story. The 2026 concessions are almost entirely the philosophical use of “on reflection” in discussions of Taylor and Rawls, plus you describing other people updating: Thiel, Joyner, Tom Holland, Skip Bayless never once saying it in twenty years.
Which is the joke sitting inside the result. Your archive contains a great deal of writing about the virtue of saying you were wrong. In October 2025 you proposed demoting any commentator who never says “I overstated this.” In December you wrote that more willingness to say “I was wrong about this specific thing” is a passport. In January 2026 you criticized people who write “mistakes were made” rather than “I made a mistake.” In February you wrote that Alliance Theory predicts apocalypse talk replaces “I was wrong” with “we were doomed.”
That is a man who has thought hard about the practice and performed it once in nineteen years.
This completes the brakes picture and it is the least flattering thing in the project. Fifty-eight posts outside published status, 0.21 percent. One public correction. The brakes that leave an artifact are the ones you don’t use. What you do instead is unpublish quietly, or say nothing, and the archive keeps the original claim until it doesn’t.
PPPPS. The silent-edit check turns up less than it first appeared.
Across 27,737 posts, 2,674 were modified a year or more after publication. Between August 1 and September 30, 2023, 1,673 posts were edited thirty or more days after publication. Twenty-eight separate days saw twenty or more posts touched, peaking at 164 on September 10 and 113 on September 4. That is a sustained six-week pass through the archive.
What it targeted is the interesting part. Of the 1,730 posts edited in the second half of 2023, essentially none come from before 2012. The distribution: 215 from 2014, 501 from 2015, 455 from 2016, 160 from 2017, 219 from 2018, 81 from 2019. You went at 2014 through 2019 and left the rest alone.
Filter to days where at most two old posts were touched, which isolates individual revisits, and you get 283 across nineteen years. Fifteen a year. Most concentrated in the 2007 to 2011 posts, and the modifications themselves cluster recently: 35 in 2026, 15 each in 2024 and 2025.
Villaraigosa’s Marriage Kaput, published January 29, 2007, edited February 24, 2026. Drunk On Kerry Howley, 2008, edited June 4, 2026. My Interview With Author Amy Bloom, 2009, edited June 27, 2026. Did I Join A Cult And Not Even Know It, 2009, edited February 2026.
luke_ford_master_interview_catalog
PPPPPS. First-person density is flat across 2001 to 2011. It runs 35.4, 33.9, 36.3, 36.1, 35.1, 38.9 through the static years, and 38.9, 32.4, 39.4, 35.1, 33.3 through the early WordPress years. Two different platforms, two different subjects, one stable habit for a decade. The seam between the corpora is invisible, which is itself a check on the method: two independently parsed archives agree at the join.
That makes 2012 and 2013 a genuine excursion rather than a peak on a rising curve. The series sits at 35 for eleven years, spikes to 52.6 and 63.3, then falls off a cliff to 19.8 and never recovers. What I called a peak was a two-year event, and it is the period of the twelve-step writing, the one-man play, and the provocation essay. Not the culmination of a tendency. A different project, briefly.
The accusation rate is the stronger correction. It runs between 17 and 36 per hundred thousand words for the entire porn era and stays in that band for most of the following two decades. 2001 is 18, the lowest year in the whole series. 2026 is 27.
So the man Wired called the most hated man in web porn was accusing people at roughly the same rate he does now writing about sociologists. Whatever made the early Ford dangerous, it was not accusation density. The peaks in the series are 2021 at 73 and 2020 at 50, the Coulter and voter-fraud years, and the porn years sit at the bottom.
That damages a claim the comprehensive essay makes on other people’s authority. It describes the early Ford through the Wired and Jewish Journal portraits, and those portraits are about publication threshold, verification and target selection. They are not about how often he called people liars, and the text says he did that less then than later.
The link data adds one thing. External links per hundred thousand words run 85 in 2001 and rise to 466 by 2005, which is the same rate range as 2026. The sourcing density you built in the AI era is the sourcing density you had in 2005.
PPPPPPS. I asked whether 2026’s subjects are new people or the same people at greater length. They are not people. The top entity in your 2026 archive is Alliance Theory, in 1,509 posts, followed by David Pinsof at 705, Ernest Becker at 520 and Stephen Turner at 378. The most-named humans are the theorists supplying the frames. Four of the five things you write about most this year did not exist in your archive before 2022, and three of them were near-zero before 2025.
The Sailer arc. Few through 2012. Twelve in 2013, 116 in 2014, 700 in 2015, 847 in 2016, then 230, 114, 15. From absent to the single most-mentioned name in your archive in two years, and back down in three.
And it lines up with the rampage. Those years, 2014 through 2019, are exactly the publication years the August and September 2023 editing targeted: 215 posts from 2014, 501 from 2015, 455 from 2016, 160 from 2017, 219 from 2018, 81 from 2019. Nothing before 2012. You went back through the Sailer period and nothing else.
[LF: I used to love right-wing comment sections, including Sailer’s. Then during Covid and afterward, I hated much of it. I thought the Covid conspiracies were unconvincing and it made me re-evaluate the crowd, and so I went back through my archive to delete most of these comments, even though they were on different topics.]
Deleting quotes of comments from other blogs from your own posts is a brake nobody has ever proposed to you and nobody can see. It costs you traffic on old pages, it produces no essay, it takes six weeks, and it leaves the reader no way to know it happened.
Chaim Amalek follows the same shape a decade earlier, running 26 to 41 mentions a year from 2007, spiking to 248 in 2015, then 109, 10, 8, and zero from 2019 onward. Alexander Technique peaks at 176 in 2011 and reaches zero in 2026.
The turnover figure gives the base rate. Between 49 and 78 percent of each year’s top 150 names are absent from the previous year’s top 150, every year, for twenty years. Your subject matter has always churned this hard. The 2026 shift is not unusual in magnitude. It is unusual in kind, because for the first time the churn produced concepts rather than people.
PPPPPPPS. The Prager corpus splits into three, and the 1998 layer is the oldest measurable Ford in existence.
The dated show notes, 41 files and 75,821 words from 1998 and 1999, run at 10.2 first-person per thousand and 79 accusation terms per hundred thousand. Both numbers are outside every other era. First-person at 10 is where you sit in 2015 and 2020, not where you sit in 2001 through 2011, which held steady at 35. And 79 accusation terms per hundred thousand is the highest figure anywhere in the archive, more than double the 2001 to 2006 range and above the 2021 peak of 73.
So the earliest surviving Ford writes like the 2020s Ford: absent from his own prose, high accusation density. Then he becomes personal for a decade, then goes back.
The explanation is genre rather than character, and that’s the point. Radio-show notes are hour-by-hour summaries of another man’s broadcast with quotes and bracketed commentary. The form suppresses the first person and raises the accusation rate because the subject is a talk host making claims about liars and hypocrites all day. Two hundred and twenty-five bracketed editorial interpolations sit in the corpus, about 49 per hundred thousand words, which is you commenting inside somebody else’s argument without stepping into the sentence.
That is the same structural move as the 2026 essays: analysis of another man’s positions, in prose from which you have removed yourself, at length. The form you arrived at last September has a 1998 ancestor.
The two long biographies are the other half: 200,730 words across two files, 37.3 first-person per thousand, revised for decades, with internal date references from 1918 to 2026. Those are the personal register, and they’re the ones you’ve kept coming back to.
The middle era, 2001 to 2013, is when the personal register took over everything. Looked at across twenty-eight years rather than twenty, that’s the excursion, and the impersonal analytical mode is what you started with.
PPPPPPPPS. The autobiography is 178,000 words across twenty-one chapters. First person runs at 47.0 per thousand across the whole thing, with chapter twenty at 90.9 and chapters two, three and seven-a above 60. That is higher than any year in the blog, including the 2013 excursion at 63.3. So the top of the self-reference range isn’t a blogging year at all. It’s the memoir, which is what a memoir should look like, and it gives the series a proper upper bound: 91 at one end, 1.7 in 2026 at the other, a fifty-three-fold spread across one man’s output.
Accusation density averages 70 per hundred thousand, which sits near the archive maximum, and the chapter variance is enormous. Chapter eighteen contains zero accusation terms across 15,100 words. Chapters four and five run 278 and 234. So the book alternates between stretches with none and stretches at three times the blog’s worst year.
