A graduate recruiter at a mid-size firm reads somewhere between four hundred and a thousand applications in a season. They are not running your cover letter through a detector. They do not need to. By the time they hit the third letter that opens "I am writing to express my enthusiasm for" and closes "I would relish the opportunity", the pattern is doing the work for them.
That is the actual problem with using AI to write applications, and it is not the one people worry about. The worry is getting caught by a tool. The reality is getting binned by a human who has read the same sentence four hundred times and has stopped seeing it.
So the question we had to answer while building tiero was narrow and mechanical. Not "is AI writing bad", but: what specifically makes generated text recognisable, and can we stop a model producing it?
It turns out somebody had already done the work.
Wikipedia had the answer, and it was 15,000 words long
Since 2023, a group of Wikipedia editors called WikiProject AI Cleanup has been finding and fixing undisclosed AI-generated content across the encyclopedia. Doing that at scale meant they had to get very good at spotting it very fast, and somewhere along the way they started writing down what they were noticing.
The result is a page called Signs of AI writing. It is roughly fifteen thousand words. It is not a detector and it does not pretend to be. The editors are explicit that it is descriptive rather than prescriptive, a field guide of observations rather than a set of rules, and that no single item on it proves anything on its own.
That caveat is the reason the list is any good. Most "how to spot AI" content is somebody's vibe dressed up as a method. This is a year of people reverting the same edits over and over, writing down what kept recurring, and arguing about it in public on the talk page.
A few of the patterns, once you see them, are impossible to unsee:
The em dash. Models reach for it constantly, and usually in positions where a comma or a bracket would be more natural. It is the single most reported tell, to the point where writers who legitimately love em dashes are now finding themselves accused.
The rule of three. Three adjectives. Three clauses. Three reasons. Generated prose falls into triples whether or not the world happens to contain three of the thing.
Negative parallelism. "It isn't just X, it's Y." "Not merely a tool, but a partner." Once you notice it you will find it in every AI-written paragraph you read for the rest of the week.
Puffery. A flat statement of fact, followed by a clause explaining that the fact was significant. Text that keeps telling you it matters instead of being something that matters.
Uniform sentence length. Less quotable than the others and probably the most reliable. Human writing lurches. It runs long, then stops. Generated writing settles into a rhythm and holds it for six hundred words.
What we did with it
We took the list and rewrote it as a constraint, aimed at first-person early-careers writing rather than encyclopedia prose. It lives in one file in our codebase, ai-tells.ts, and it gets appended to the prompt of every generator tiero runs: application answers, draft answers, cover letters.
Some of it is a straight banned-words list, adapted for the words that actually show up in graduate applications rather than the ones that show up in Wikipedia articles. Passionate. Thrilled. Delve. Leverage. Robust. Seamless. Proven track record. Wealth of experience.
The more interesting rules are structural:
Use plain verbs. "I was a shift supervisor", not "I served as a shift supervisor".
Never bolt an -ing clause onto the end of a sentence to comment on what you just said. Not "which cut waiting times, demonstrating my ability to problem-solve". Stop at the fact. The reader can draw the conclusion; that is what reading is.
Repeat a noun instead of cycling synonyms. "The company" stays "the company". It does not become "the firm", then "the organisation", then "the business" across four sentences.
Vary sentence length hard. A four-word sentence next to a twenty-five-word one.
Never write a sentence whose only job is to say that something mattered.
And, yes: no em dashes. Which is a slightly absurd rule to enforce on a machine on behalf of a punctuation mark that did nothing wrong. But the tell is the tell.
The rule we got wrong first
The first version of this was too aggressive, and it produced something worse than what it replaced.
Strip enough out of a piece of writing and you do not arrive at natural. You arrive at a different affectation: clipped fragments, self-consciously plain vocabulary, a sort of performed bluntness that reads like someone doing an impression of honesty. It has its own rhythm, and that rhythm is now its own tell.
So the last paragraph of the file is a warning against the rest of it. The target is a real person writing carefully, not a stylised anti-AI voice. Contractions are fine. Ordinary connective words are fine. The instruction is to say the true, specific thing, which is a much harder problem than avoiding a word list, and it is the one that actually determines whether a letter is any good.
This is also where the real limit sits. No prompt makes writing worth reading. The tells are a floor, not a ceiling. What actually distinguishes a graduate cover letter is having something specific in it, which is why most of tiero is not a writing tool at all but a system for getting the specific things out of your head and somewhere you can find them again. A letter built from a bad shift you worked at a hotel will beat a beautifully de-slopped letter built from nothing.
Why we put the source in the product
Underneath generated answers in tiero there is a line that says "Written against the known signs of AI writing". Click it and you get a short explanation and a link to the Wikipedia page.
That is deliberate, and it says something narrower than you might expect. It names the standard we write against. It does not claim your text will pass an AI detector, because detectors are unreliable in both directions and false accusations land on real students. That is not a promise anyone can honestly make, so we do not make it.
What we can say is what we are aiming at, and point at the public document we are aiming at it with, so you can go and check whether we are any good at it. That felt like the minimum, given the whole product is an argument about not taking generated text at face value.
Where it shows up
The constraints run on every generator. The Cover Letter Coach also runs the list in the other direction, reviewing a finished letter and flagging the places it still reads as generated, including ones you wrote yourself. Most people are startled by how much of their own unassisted writing trips it. "Passionate about" was there long before ChatGPT was.
Application Questions uses the same constraints when drafting answers from your Experience Library.
If you want to read the source list, it is here, free, and worth an hour of anyone's time whether or not they ever use an AI writing tool. It is one of the more useful things on the internet about how to write, and it got there by accident, as a side effect of a few hundred volunteers cleaning up after everyone else.
Written by
Adam Hannigan
Founder, tiero
Published
