On GitHub, the vocabulary everyone blames on language models is still climbing. I ran the same word list over 25,000 arXiv abstracts expecting the same line. Science got there a year earlier, peaked in early 2025, and has been handing the words back ever since.
The law telling federal agencies to write clearly turned sixteen last month. Thirty years of the Federal Register show the government's prose did get plainer — a decade before the Act, on the strength of a memo with a deadline, and not at all since.
Scientists are trained to qualify their claims. Across 2,400 arXiv abstracts, how much a field hedges turns out to track something simpler than caution — whether it observes the world or proves things about it. And despite the reputation, hedging isn't on the rise.
Codeberg's members just voted to ban vibe-coded projects, which is one way of saying the question has stopped being rhetorical. I build a prose linter, so I pointed it at a decade of GitHub instead — and the load turns out to be real, to land on pull requests rather than issues, and to look nothing like the tells everyone repeats.
At Supreme Court oral argument the justices share a fixed hour, and they split it about as unevenly as they split anything else. Six terms of transcripts show who spends the Court's time, who barely does, and the newest justice who out-talks them all.
The NBA grades its own officiating in the final two minutes of close games and publishes the results. Eighty-nine thousand graded calls later, the errors have a shape — and the question everyone wants answered turns out to be the one the data is worst at.
Wikipedia has no editor-in-chief; disputes end when someone hits undo. Read the last five hundred edits of its own most-contested articles and the map of where the reverting runs hottest is not the one the culture wars would draw.
A library for evaluating well-formed formulas, built to find out how much of an introductory logic course automates cleanly. Written in 2016; the demos here run the real parser.