Who Trained Us to Say "Unalive"?

On 8 April 2022 the Washington Post put a name to something users of large platforms had been doing for some time: algospeak. People wrote "unalive" instead of dead and "SA" instead of sexual assault, and coined phrases like "le dollar bean", because the systems that rank and remove posts punished the plain words. No law required it and no editor asked for it. A classifier did, simply by deciding what got seen. The move is older than the name: on the bulletin boards of the 1980s, users wrote "h4x0r" and "w4r3z" partly to slip past sysops' keyword filters, and leetspeak was the result. Three years later the same pressure arrived from another direction. Through 2025, writers reported dropping the em dash, a mark English prose has leaned on for two centuries, because readers had come to treat it as a sign of a chatbot; The Ringer ran the story in August 2025. Teachers now pass essays through AI detectors, and editors and hiring managers do the same with pitches and cover letters. The incentive for a human writer has become plain: write in a way the judge will not flag.

A measurement lab recognises this pattern. When a measure becomes a target, it ceases to be a good measure: the anthropologist Marilyn Strathern's phrasing of Goodhart's law. A text classifier was built to describe writing. Once people know it is watching, they write to it, and it starts shaping the thing it was meant to measure. What is new is that the judges now train both sides. Language models are tuned toward prose that reviewers find safe and polished. Human writers are being tuned away from whatever reads as machine-made. Both drift toward the narrow band the judges leave alone. There is published evidence of narrowing already: Doshi and Hauser (Science Advances, 2024) found that AI assistance raised the quality of individual stories while making the stories more alike. The costs of being misjudged are not shared evenly either: Liang and colleagues (Patterns, 2023) found widely used detectors flagging essays by non-native English writers as AI-written at high rates. The lesson we take is the one our own instruments are built on. Judge the work by what was done, not by how it sounds. A distinct voice is not evidence of anything, and a voice that has been sanded down to pass a filter is not evidence of honesty. Turing asked whether a machine could pass as a human. The question has quietly turned around. For twenty years we have clicked on traffic lights to prove to software that we are people; CAPTCHA stands for Completely Automated Public Turing test to tell Computers and Humans Apart. Prose is now becoming that test, a reverse Turing test that nobody agreed to take, and the writers most likely to fail it are those whose natural style already reads as plain or clean, non-native speakers first among them.

Three things. First, whether institutions keep judging writing by its style. Turnitin switched on its AI detector for universities in April 2023; by August, Vanderbilt had switched it off, reasoning that even the vendor's claimed 1% false-positive rate meant roughly 750 of its 75,000 annual papers wrongly flagged. Others followed. The alternative gaining ground is provenance: a record of how a document was produced, rather than a guess from its surface. Second, consolidation. In June 2026 the largest writing-assistant company agreed to buy one of the two best-known AI detectors, so the tool that helps you write and the tool that judges whether you did now sit under one roof. How that company handles the conflict is worth following. Third, a small and cheap indicator: whether the em dash comes back. If writers start using it again without apology, the pressure to write for the judge is easing. If it keeps disappearing, people are still editing their own voices to satisfy a machine, and the band of acceptable prose is getting narrower for everyone in it.