AI Writing Tell Checker
This is not a probability-guessing detector. It checks your text against a curated list of 335 vocabulary tells and seven structural patterns that human editors flag in machine writing, and shows you exactly where each one is.
per 1,000 words
Annotated text
Vocabulary tells
Structural tells
What this checker looks for
The wordlist comes from an editorial banlist built over hundreds of humanization passes on LLM-drafted articles: 335 entries across eight groups, from fingerprint vocabulary (delve, tapestry, multifaceted) through stock openers, hedging structures, and the pseudo-casual reveal patterns ("here's the kicker") that models reach for when told to sound conversational. On top of the wordlist, seven structural checks catch what single words cannot: the "it's not X, it's Y" contrast tic, neat A-B-and-C triads, staccato fragment chains, em-dash density, paragraphs opening with And or But, rhetorical question stacking, and uniform sentence length.
How to use the results
Work the highlights, not the score. Each highlighted span is a rewrite candidate: swap the fingerprint word for the plain one, cut the filler phrase, break the triad's rhythm by making one item longer or moving it. Structural flags point at sentences to recast rather than words to swap. When the per-1,000-word score drops under 5 you are inside the range of normal human writing; most raw LLM output starts between 25 and 60.
The honest limits
A wordlist cannot see token-level statistics, and a clever writer can produce slop that contains none of these patterns. What it does reliably is catch the patterns readers and editors actually notice, which is the standard that matters when a human decides whether to keep reading. Pair it with the sentence-variance check for rhythm analysis.
Frequently asked questions
How is this different from AI detectors like GPTZero?
Detectors output a probability score from a statistical model, and they are wrong often, in both directions. This tool does something humbler and more useful: it checks for specific, named patterns that editors recognize in LLM output, and shows you each occurrence so you can rewrite it. You get an edit list, not a verdict.
If I fix every flag, will my text pass AI detectors?
It will read more human, which is the part that matters for readers and for editors. Detector scores are influenced by deeper statistical properties like token predictability that no wordlist can fully control. Fixing tells removes the patterns humans notice first, and in our experience it substantially shifts detector scores too, but no tool can promise a pass.
Are these words banned from good writing?
No single word is. Humans write 'framework' and 'robust' too. The signal is density: machine-generated text stacks these patterns far more often than a human editor would tolerate. That is why the score is per 1,000 words and weighted, not a simple yes or no on any word.
Is my text uploaded anywhere?
No. The full analysis runs in your browser with JavaScript. Nothing you paste leaves your machine, and there is no server, no logging, and no word limit.