Em dashes in AI writing: we measured our own copy, found the rule had never worked, and removed 23,200 of them
FCT's writing standard banned eight AI tells and missed the one readers notice. We measured our copy, found an em dash every 50 words, and took 23,200 out.
# Em dashes in AI writing: we measured our own copy, found the rule had never worked, and removed 23,200 of them FCT's writing standard banned eight AI tells and missed the one readers notice. We measured our copy, found an em dash every 50 words, and took 23,200 out. By Fernando Caudillo Tafoya · FCT Technologies · August 14, 2026 You can usually tell. You land on a company's page and something about the sentences reads machine made. You could not point at the word that did it, and you already trust the page a little less. That cost lands before anyone reads the offer. FCT writes with AI in the loop, so this is our problem before it is anyone else's. We have kept a house writing standard since the company started. One of its rules is a list of AI tells not to write: no significance inflation, no rule-of-three padding, no "let's dive in," no generic upbeat closer. Eight of them, by name. On August 13 we stopped trusting the list and measured what we were producing. 19.7 em dashes per 1,000 words. That is one every fifty words, against a standing rule that said avoid them. Beside it, 17.3 bold spans per 1,000 words, against a rule that said no reflexive boldface. Neither rule had ever once operated. Then the part that costs money. The voice guide that governs every word of client copy, every service page and every proposal listed those eight tells and did not list the em dash. That is the single most recognizable machine-writing marker there is. So every piece of copy that guide produced arrived carrying the tell, however often we cleaned the output afterward. We took it out: about 23,200 marks across some 640 files, and then the storefront itself. As of today the homepage, the service pages, the newsroom and `llms.txt` all return zero. You can check that rather than take our word for it, which is the only kind of claim worth making about your own writing. Two things nearly let the whole thing pass while reading clean. Our own detector was pointed at a shallow file pattern and saw 2 of 112 files in one area, so about 1,890 marks were invisible while that area reported healthy. And on the website, the service and project pages render from their component rather than from their markdown body, so a markdown-only sweep would have left every page a visitor sees still carrying the habit. Both failed toward clean, which is the direction nobody catches. What we left alone is the other half of the standard. Anything a person said stays exactly as they said it. So does copy already published where our version is no longer the live one, like a posted update or a submitted directory listing: editing our copy only puts it out of step with a record nobody can change. The contact form's email subject kept its punctuation too, because mail routing matches on it. It runs as a standing monthly check now. It measures the rate per area against a ceiling and reports without editing anything. The ceilings sit above zero on purpose, since a quotation or a heading legitimately carries the mark, and a check that can never come back clean is one people learn to ignore. Availability, plainly: the rule went into the guide that governs client work on August 13, the pipeline that writes our own content now writes against that guide instead of against nothing, and the site is deployed. This is not a product and we are not selling it as a service. The limit is worth saying too. Removing a tell does not make writing good. It removes the thing that argues against you before the reader gets to the argument. The rest is whether the sentence says something true and specific, and no checker can measure that. If we are going to audit your site for the things that quietly cost you trust, ours has to survive the same audit first. The rule underneath it is the one we would give away for free: a style rule nobody measures is one nobody keeps.You can usually tell. You land on a company’s page and something about the sentences reads machine made. You could not point at the word that did it, and you already trust the page a little less.
That cost lands before anyone reads the offer.
FCT writes with AI in the loop, so this is our problem before it is anyone else’s. We have kept a house writing standard since the company started. One of its rules is a list of AI tells not to write: no significance inflation, no rule-of-three padding, no “let’s dive in,” no generic upbeat closer. Eight of them, by name. On August 13 we stopped trusting the list and measured what we were producing.
19.7 em dashes per 1,000 words. That is one every fifty words, against a standing rule that said avoid them. Beside it, 17.3 bold spans per 1,000 words, against a rule that said no reflexive boldface. Neither rule had ever once operated.
Then the part that costs money. The voice guide that governs every word of client copy, every service page and every proposal listed those eight tells and did not list the em dash. That is the single most recognizable machine-writing marker there is. So every piece of copy that guide produced arrived carrying the tell, however often we cleaned the output afterward.
We took it out: about 23,200 marks across some 640 files, and then the storefront itself. As of today the homepage, the service pages, the newsroom and llms.txt all return zero. You can check that rather than take our word for it, which is the only kind of claim worth making about your own writing.
Two things nearly let the whole thing pass while reading clean. Our own detector was pointed at a shallow file pattern and saw 2 of 112 files in one area, so about 1,890 marks were invisible while that area reported healthy. And on the website, the service and project pages render from their component rather than from their markdown body, so a markdown-only sweep would have left every page a visitor sees still carrying the habit. Both failed toward clean, which is the direction nobody catches.
What we left alone is the other half of the standard. Anything a person said stays exactly as they said it. So does copy already published where our version is no longer the live one, like a posted update or a submitted directory listing: editing our copy only puts it out of step with a record nobody can change. The contact form’s email subject kept its punctuation too, because mail routing matches on it.
It runs as a standing monthly check now. It measures the rate per area against a ceiling and reports without editing anything. The ceilings sit above zero on purpose, since a quotation or a heading legitimately carries the mark, and a check that can never come back clean is one people learn to ignore.
Availability, plainly: the rule went into the guide that governs client work on August 13, the pipeline that writes our own content now writes against that guide instead of against nothing, and the site is deployed. This is not a product and we are not selling it as a service.
The limit is worth saying too. Removing a tell does not make writing good. It removes the thing that argues against you before the reader gets to the argument. The rest is whether the sentence says something true and specific, and no checker can measure that.
If we are going to audit your site for the things that quietly cost you trust, ours has to survive the same audit first. The rule underneath it is the one we would give away for free: a style rule nobody measures is one nobody keeps.
Frequently asked questions
Why does an em dash matter to a business website?
It is the most recognizable marker of machine-written prose, and a reader registers it without being able to name it. The cost is not stylistic. A page that reads generated is trusted less before anyone reaches the offer, which argues against whatever the page claims. Buyers see more generated copy every month, so the tells get easier to spot, not harder.
Can't you just find and replace it?
No, and that is the main finding. The mark does a different job in nearly every sentence, and each mechanical swap looks fine on its own while the paragraph goes flat. Only about a quarter of instances were safe to swap by machine. Roughly 40% in dense documents were a label introducing its own definition, which wants a colon and reads as a shopping list under a comma. We judged each one, with tools that refuse a whole batch if any line moved underneath them.
How do you know it stays fixed?
A standing monthly check measures the rate per area against a ceiling and reports without editing anything. The ceilings sit above zero on purpose, because a quotation or a heading legitimately carries the mark, and a check that can never come back clean is one people learn to ignore. The rule also lives in the guide that governs client copy, so it applies when we write, not as a cleanup pass.
Is this something FCT sells?
No. There is no product or service here. It is the writing standard behind the work we already do, applied to our own storefront first. What transfers to your site is the rule underneath it: a style rule nobody measures is one nobody keeps.