Edited by us
AI writing tells — where AI gives itself away
Not in the grammar — in the habits. Seven tells, how to remove them, and the overcorrection trap.
By Mephify (operator)3 min read
The tell is habit, not grammar
Flawless grammar, formal vocabulary, a single "however" — none of these is a tell. People write that way too. The tell is clustering: when three or four of the habits below land in one piece, readers know, even when they cannot say why.
- Inflated vocabulary: "delve", "tapestry", "testament", "pivotal", "landscape", none of them doing actual work in the sentence.
- The "not just X, but Y" reflex — and everything arriving in threes, whether or not there are three things to say.
- Em-dashes on every line, bold on every third phrase, a heading over every short paragraph.
- Signposting: "Let's dive in", "In this article, we will explore". Then staged staccato: "One goal. Zero compromises."
- Elegant variation: the same thing renamed each sentence to dodge repeating a word, until nobody is sure it is still the same thing.
- Stock openers and closers: "Great question!", "In conclusion", "The future looks bright."
- Hedge stacks: "could potentially, in some cases, depending on context". Three qualifiers guarding one small claim.
The overcorrection trap
Trying to erase every signal produces the opposite tell. One em-dash, one "however", clean grammar — that is just writing, and human writing has all of it. Fix clusters only, and never touch quotations, titles, proper nouns, or anything inside code. Renaming every repeated word into a synonym is itself the machine habit you were trying to remove.
The most dangerous fix is the one that sounds safest: telling the AI only to "be more specific." A demand for specifics must travel with its pair: "use only facts from the material I gave you and from checkable sources." Without the pair, the model invents dates, numbers and little anecdotes to sound human. A natural-sounding fabrication is worse than an awkward truth.
How to put it in a request
Add these lines to the end of a request
Vary sentence length and structure. Skip stock openers and wrap-ups. No 'In conclusion'. End on a concrete fact. Be specific, but use only facts from the material I gave you. Invent nothing.
Three sentences, but they block most of the list above. The one that matters most is the last: the pairing rule.
- 1Read it aloud — does every sentence land with the same rhythm and length?
- 2Is there a stock opener or closer — "Great question", "In conclusion"?
- 3Are bold text and bullet lists carrying content, or standing in for it?
- 4Any abstract praise words with no how or what behind them?
- 5Start with any number or anecdote you could not source if asked.
How Mephify handles this
Instructions made here carry these style rules and a pre-submission self-check from the start, the fact-pairing rule included, so you do not have to re-attach those lines to every request. The rules trace back to the "Signs of AI writing" page kept by Wikipedia's editors and a public humanizer ruleset, with a separate rule set maintained for the habits of Korean-language writing.
Reading helps; trying is faster.