AI tell · grade B
An AI writing style guide, graded entry by entry
Most guides to AI writing patterns are prose, a list of things to avoid, argued from the writer's ear. underslop runs an actual checklist instead, the same one behind every scored tell page on this site. Each entry carries a grade: E for a pattern grounded only in our own model logs, B where those logs line up with outside corroboration, S for a pattern our social-sentiment review surfaced and confirmed, and L for a pattern at least two independent style guides already agreed on. Roughly twenty entries clear the bar to count toward a score. A separate set gets tracked and shown but never added to that score: greylist words that read fine in small doses, an isolated em dash, a curly quote, raw adverb density, and the handful of once-popular rules our own review found didn't survive scrutiny. That second list is the part most guides skip, because admitting a rule doesn't hold up costs more than quietly repeating it.
What it is
The checklist behind this page lives in our own build, generated from our model logs, our social-sentiment review, and three outside style guides cross-checked against both. Every entry records where it came from: an empirical log reference, a corroborating guide, a social citation, or some combination. A pattern only earns a score if it clears two bars at once, real evidence backing it and no later finding that contradicts it. A pattern that fails either bar still gets computed on every draft. It shows up in a separate report. It just never moves the number a writer is trying to bring down.
How often it shows up
As of this checklist's current version (research/eval/tell-checklist.json), twenty distinct patterns are graded and scored: four vocabulary words, three stock opening phrases, seven sentence-level structures, four chatbot-artifact strings, and two formatting rules checked line by line rather than by regex. Eight more categories are graded but excluded on purpose, moved out after our social-sentiment review found the popular version of each rule didn't hold up against real writing. That split, twenty scored against eight knowingly set aside, is the actual shape of the evidence, not a guess at what sounds like AI.
Why readers flag it
A prose guide asks a reader to trust the writer's ear. A graded checklist lets a reader check the writer's claim: run the same text through the same pattern and get the same answer regardless of who is arguing. That matters most exactly where a guide is most tempted to overreach, listing every popular tell as equally certain when some of them turned out to be common in ordinary human writing once anyone actually checked. A grade is the honest version of a confidence interval. Most style guides don't publish one.
The fix
Use the grade before you use the rule. A scored E or B pattern is worth fixing on sight: it is backed by real logs, by outside corroboration, or both. A pattern that only shows up in the report list, adverb density, a greylist word, an isolated em dash, is worth a glance and rarely worth a rewrite on its own. Treat the report list as a second opinion, not a verdict, and spend the actual editing time on what the score is built from.
Examples
You're absolutely right that clarity matters. This isn't just about grammar, it's about respecting your reader. Let me know if you'd like me to expand on any of these best practices!
You're right that clarity matters. This isn't just about grammar—it's about respecting your reader.
In today's fast-paced world, a strong content strategy isn't a luxury, it's a necessity. Our approach doesn't just deliver results, it delivers a testament to what's possible when a brand truly commits.
A strong content strategy is a necessity now, not a luxury. When a brand really commits, the results show what's possible.
Questions
How do I write a style guide for avoiding AI patterns?
Start from evidence, not instinct. List the patterns you can actually test: a regex, a word count, a windowed density, and grade each one by where it came from, your own samples, an outside source, or both. Then track the patterns you suspect but cannot confirm separately, so the guide says plainly which rules are backed by real evidence and which are guesses.
How do I avoid AI text patterns?
Fix what the checklist actually scores first: a validation opener before the real content, and a claim propped up by a source that is never named. A paragraph turned into bold-label bullets when it should have stayed prose is worth checking too. That covers most of what a reader consciously notices. Everything in the contested list, a stray em dash, a greylist word, is worth a glance but rarely worth a rewrite by itself.
What are the prompts to avoid typical AI writing patterns?
None that we have tested and trust. A prompt is an instruction to the model before it writes; a checklist is a check on the text after it exists. We grade the second because we can verify it against real output. Whether a prompt reliably prevents these patterns in the first place is a separate question we have not measured, and we would rather say that plainly than promise something we have not tested.
How do I spot AI writing patterns?
Read for the shape before the words. A reply that opens by praising the question and a paragraph that turns into bullets for no reason are both shape problems, not vocabulary problems. Those shapes repeat across models more reliably than any single word does, which is why the checklist scores structure and format alongside vocabulary instead of treating a banned-word list as the whole job.
What are overused words in AI writing?
A short list of vocabulary items are graded and scored on this site: inflated-metaphor nouns, hype adjectives for launches, and stock scene-setting openers like the fast-paced-world line. A much longer list is graded but only reported, greylist words that read fine in moderation and are common in ordinary professional writing too. The scored list is short on purpose. Most of what makes writing sound generated is structure, not vocabulary.
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.