I gave Claude my entire WordPress archive.<\/p>\n
Claude says (and I have edited down its prolixity): The full WordPress export of lukeford.net makes a kind of study possible that could not be run on samples. It also makes a demonstration possible that is more useful than the study, because for once the person interpreting the numbers wrote the corpus and knows what he meant.<\/p>\n
Here is the setup. Take a standard accusation vocabulary of the sort used in computational content analysis: liar, lying, lied, fraud, fraudulent, dishonest, dishonesty, deceptive, deceit, hypocrite, hypocrisy, coward, plagiarism, shill, grifter, charlatan. Search all 27,737 published posts in Luke Ford’s (b. 1966) archive, 2006 through August 2026. Search author prose only, stripping every blockquote first, so nothing quoted in block form counts. Require the sentence to also contain a proper-name proxy, meaning two consecutive capitalized words, on the theory that an accusation needs a target. The search returns 1,519 sentences.<\/p>\n
Of those, 516 fall in 2026, a year that is seven and a half months old. The next highest year is 2016 with 202. The five years an earlier essay in this series identified as Ford’s most combative, 2008, 2015, 2016, 2020 and 2021, together produce 570, barely more than 2026 alone.<\/p>\n
The headline writes itself. Luke Ford used accusatory language more in the first two-thirds of 2026 than in any full year of his blogging life.<\/p>\n
That sentence is true as measured and false as understood, and the correction requires no judgment at all. It requires a denominator.<\/p>\n
Ford published 11,421,248 author-written words in 2026 through August 19. Divide the hits by the words and 2026 returns 4.5 accusation-vocabulary sentences per hundred thousand words. The archive average across twenty years is 6.3. The peak is 2021 at 15.3, followed by 2020 at 12.9, 2024 at 10.7 and 2014 at 10.3. On this measure 2026 is the fourth-lowest year in the entire archive, below 2008, below 2016, below every year of the alt-right period, and less than a third of 2021.<\/p>\n
Same corpus. Same dictionary. Same search, run once. Two normalizations, both defensible, opposite conclusions, and no coder ever touched it. Anyone who has ever been reassured that a method is objective because it involves no human judgment should sit with that for a moment.<\/p>\n
The rate series has an additional property worth noting, because it converges with a reading arrived at independently. The longitudinal essay in this series<\/A> argued that Ford’s 2020 self-regulation rules produced no immediate behavioral discontinuity, and that the 2021 archive shows better sourcing wrapped around the old move from proposition to person: the voter-fraud post declaring that believers are not clear thinkers, the Ann Coulter post<\/A> concluding that she is lying, the dentistry opening. That reading came from reading. The word-rate series, which knows nothing about any of it, puts 2020 and 2021 at the top. The qualitative account and the normalized count agree with each other and both disagree with the raw count.<\/p>\n Now the composition, which is where a dictionary earns its reputation.<\/p>\n Of the 516 hits in 2026, 220, or forty-three percent, contain fraud or plagiarism used as subject matter rather than as an accusation Ford is making. Ponzi fraud in a study of affinity networks. Anti-fraud software in a biography. The Justice Department’s civil-rights fraud posture. A seventeenth-century pamphlet whose title translates as new criticism or new fraud. Fraud is the single most common hit in the whole archive, 175 times in 2026 and 219 times across the five comparison years, and it is doing the work of an ordinary noun almost every time.<\/p>\n Another 163 of the 516, thirty-two percent, contain quotation marks, most of them reproducing somebody else’s words inline rather than in a block. Eight are self-directed, Ford calling himself a coward or describing his own dishonesty, which a hostility measure scores as aggression and which is nearer its opposite. Four are the grammatical homonym, the verb that means to recline rather than to deceive, as in a sentence about lying back doing nothing.<\/p>\n The revealing part is that these proportions barely move across twenty years. In the five comparison years the topic share is forty-one percent against 2026’s forty-three, and the quotation share is thirty-nine percent against thirty-two. The dictionary is not newly broken by AI-era writing. It was always this broken. What changed is the volume of prose passing through it, and volume is exactly what the raw count fails to control for.<\/p>\n One more thing belongs here, because leaving it out would make this essay guilty of what it describes.<\/p>\n Before running any of these checks, the working hypothesis was that the 516 hits in 2026 would concentrate in a handful of essays about other people’s accusations, the studies of Philip N. Cohen<\/A>, Nathan Cofnas<\/A> and Amy Wax<\/A>, where Ford writes “lies a lot” and “dishonest” dozens of times while arguing that somebody else should not have written them. That was a plausible story. It fit the year. It flattered the analysis.<\/p>\n The check contradicted it. The 516 hits come from 279 distinct posts. The top ten posts supply twenty-nine percent of them, so the distribution is dispersed rather than concentrated. The single largest contributor is not an essay about accusation at all; it is a book-length treatment of Jewish affinity networks in Los Angeles [LF: Within the context that the affinity that powers fraud in Jewish life is the byproduct of affinity generosity, which produces goodness at a massive multiple of the fraud], which produces 48 hits because financial fraud is its subject. Cohen’s study contributes fifteen.<\/p>\n So the meta-accusation family exists and is not the story. The story is topic drift. When a writer’s subjects move toward institutions, misconduct, credentialing and epistemics, an accusation dictionary lights up without a single accusation being made, and it does so in proportion to how much he writes about those things.<\/p>\n This generalizes past one blogger, which is the reason to publish it. Dictionary and classifier methods now sit underneath a great deal of published work: toxicity scores, incivility indices, sentiment panels, moderation systems, and the increasingly common practice of asking a language model to rate a corpus for hostility. Every one of them faces the four families above. Most papers that use them acknowledge the problem in a limitations paragraph and move on, because the researcher studying somebody else’s corpus usually cannot establish ground truth. He can report inter-coder agreement on a sample and hope.<\/p>\n The case here is unusual only in that ground truth is available. Ford wrote the 24 million words. He knows what he meant by them. He can state with authority that 516 hits in 2026 correspond to a number of actual accusations that is closer to zero than to 516, and no external researcher studying his archive could make that claim with the same standing.<\/p>\n What the method cannot solve is the fourth family, and it is worth being clear that no amount of dictionary refinement will. When Ford quotes Cohen calling a colleague evil in order to argue that Cohen should not have, the words on the page are identical to the words that would appear if Ford were calling the man evil. Stripping quotations removes the block quotes and leaves the inline ones. Adding a rule about attribution verbs catches some and creates new errors elsewhere. The thing that distinguishes committing an act from condemning it is the argument surrounding the sentence, which is to say it is meaning, which is to say a person has to read it.<\/p>\n Put the two headlines side by side and the front-page test does the rest. “Luke Ford used accusatory language more in 2026 than in any prior year of his archive.” “Luke Ford’s rate of accusatory language in 2026 was the fourth-lowest in twenty years.” Both are produced by the same objective method from the same complete data. A hostile writer takes the first, a friendly writer takes the second, and neither has to lie. The choice happens in the denominator, which is the part nobody reports and the part that decides the answer.<\/p>\n A final honesty note, which is also a commitment. Everything above is about an instrument. It says nothing whatever about whether Luke Ford’s accusations were warranted, which is the question that actually matters and the one this essay does not touch. That study requires reading all 1,519 sentences, marking which are in Ford’s own voice, identifying the target, and scoring each surviving accusation for demonstrated error, evidence of the target’s knowledge, and evidence of intent to deceive, on the same three-stage instrument Ford applied to Philip Cohen. It requires labeling which accusations rest on the same underlying documents, because Cohen’s record collapsed when repeated accusations proved to depend on a single email trail, and Ford’s may do the same.<\/p>\n That study has not been run. This one was easy and satisfying and let its author write about the accusation problem without reading 1,519 sentences about himself, and the temptation to let it stand in for the harder work is the exact failure his recent essays have been describing in other people.<\/p>\n The coding workbook is published, seeded and ready. The reading has not started.<\/p>\n","protected":false},"excerpt":{"rendered":" I gave Claude my entire WordPress archive. Claude says (and I have edited down its prolixity): The full WordPress export of lukeford.net makes a kind of study possible that could not be run on samples. It also makes a demonstration … Continue reading