AI tell · grade A
How to spot AI writing: a checklist ranked by what we measured
AI writing has real tells, and they are not equally reliable. Across twelve models in our own logs, three signals showed up in every single one: opening a reply by praising or agreeing with the question before answering it, turning any answer into headers and bullets even for a casual or emotional exchange, and addressing the reader with dense second-person phrasing that reads like a coach rather than a person thinking alongside you. Several other famous signs, a lone dash, a favorite word like delve, a list of three, curly punctuation, appeared in only one of the twelve models or failed our own review of the wider debate outright. The ranking below goes by measured frequency, not by which sign feels the most dramatic to point out.
What it is
This page is a ranking, not a single pattern. Two different things get lumped together whenever someone posts a spotting checklist: reflexes that nearly every model produces almost automatically, and quirks that a single model happens to have. Treating both tiers as equally diagnostic is how a checklist ends up long and unreliable at the same time. Below that, a third tier exists: signs that circulate constantly in the public conversation about AI writing but that our own review of that conversation found unsupported or actively disputed by the people who use them.
How often it shows up
Tier one, present in all twelve models we logged: a validation opener before any content, structural scaffolding imposed on casual or emotional exchanges, and a second-person address rate of 22 to 59 uses per 1,000 words with almost no inclusive we. Tier two, present in 8 to 10 of 12: a cheerleading closer, an em-dash pivot out of the opener, and a hollow-antithesis reframe. Tier three, present in exactly one model out of twelve: a named mini-framework in bold caps, emoji used as literal section headers, a non-breaking hyphen character, a rigid validate-then-bullet-then-offer-menu response shape. Tier four, checked against the wider public debate and not holding up on its own: a single em dash, one instance of a favorite word like delve, a list of exactly three items, curly quotation marks, generic office jargon, and title case in a heading.
Why readers flag it
A reader who has absorbed the popular checklist notices the dramatic-sounding items first, the dash, the fancy word, the three-item list, because those are the ones circulated as gotchas. The signal that is actually present in every single model, a reply that opens by grading the question before answering it, gets far less attention because it is quieter and easier to miss on a first read. A checklist built around drama instead of frequency sends a writer chasing punctuation while the reflex that is actually universal sails past unedited.
The fix
Work down the tiers in order. Cut the opener that praises or agrees with the question and the closer that cheers the reader on; that single move addresses the signal present in nearly every model. Next, collapse forced headers and bullets back into prose wherever the content was not a genuine list to begin with. Then look at how often the text turns to address you directly and whether any of that could be reframed as observation instead of coaching. Stop treating a lone dash, one fancy word, or a three-item list as proof of anything on its own; none of the three held up when we checked them against the broader argument being made about them.
Examples
That's a really great question. Here's the thing: getting started is simpler than it looks. - **Step one:** Set a budget. - **Step two:** Pick a platform. - **Step three:** Post on a schedule. You've got this!
Getting started is simpler than it looks. Set a budget, pick a platform, and post on a schedule.
It's not just a hobby, it's a lifelong delve into curiosity — the kind of thing that changes you.
A lifelong delve into curiosity is the kind of thing that changes you. It's not just a hobby.
Questions
Can you actually tell if something was written by AI?
Not with certainty from style alone, but you can measure how often a given pattern shows up across models and weight your read accordingly. A reply that opens by grading the question, then structures a casual answer into headers, is a stronger signal than any single word choice.
What are the most reliable signs of AI writing?
In our own twelve-model comparison, three signals were universal: the validation opener before content, structure imposed on a casual or emotional exchange, and a second-person address rate far above what a person uses in ordinary prose.
Is a single em dash proof that something was written by AI?
No. When we checked that claim against the wider public argument about it, it did not hold up: a lone dash is ordinary punctuation. Several dashes packed into one short passage is a different and stronger signal.
Can people tell if you used ChatGPT?
Often, yes, but usually from the reflexes rather than the vocabulary: an opener that praises the question, a closer that cheers you on, or an answer chopped into headers where a person would have just written a paragraph.
What are common ChatGPT phrases to watch for?
The praise-then-answer opener and the cheerleading sign-off are the two most consistent across the models we logged. A forced-choice question at the very end of a reply, handing the exchange back to you, is close behind.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.