The words AI can't stop using.
A curated selection of the AI tells we track — mined from dated corpora of real model output, generation by generation, and pruned as models change. Not a listicle: real data from our detection engine.
41 words that read machine-written in any register.
Fine words individually — people use every one of them. The tell is density: AI reaches for them constantly, everywhere, in every kind of text.
Stock phrasing that pushes a score up.
- “a marathon, not a sprint”
- “a myriad of”
- “a symphony of”
- “at its core”
- “at the end of the day”
- “beacon of”
- “bespoke”
- “by and large”
- “cacophony”
- “cognizant”
- “cutting edge”
- “cutting-edge”
- “delve”
- “double down on”
- “embark on”
- “endless possibilities”
- “ever-changing”
- “ever-evolving”
- “foster a”
- “furthermore”
- “game changer”
- “game-changer”
- “groundbreaking”
- “harness the power”
- “here's the thing”
- “highlights the importance”
- “hitherto”
- “holistic approach”
- “in conclusion”
- “in essence”
- “in the realm of”
- “in this article”
- “in today's fast-paced”
- “in today's world”
- “indelible”
- “inexorably”
- “innovative solution”
- “is not just about”
- “isn't just about”
- “it goes without saying”
- “it is important to note”
- “it is not just”
- “it is worth noting”
- “it's important to note”
- “it's not just”
- “it's worth noting”
- “kaleidoscope”
- “key takeaway”
- “lean into”
- “let's break it down”
- “let's dive”
- “let's explore”
- “let's face it”
- “let's unpack”
- “make no mistake”
- “meticulously”
- “more often than not”
- “more than just”
- “moreover”
- “moving forward,”
Mined, dated, and pruned — not vibes.
We keep dated corpora of default-voice output from each major model generation (same fixed prompts every round, so any change is voice, not topic), mine what shifted, and a human curates the list — words only join if they read machine-written in any register. When a model stops using a word, it gets demoted, not deleted: text written last year is still being pasted today. Full protocol in the methodology.
The harvest rules
- Dated corpora per model generation — same fixed prompts every round
- A human curates; nothing joins the list automatically
- A word must read machine-written in any register
- Fading words are demoted, never deleted — old AI text still circulates
About the AI vocabulary.
Why does ChatGPT say 'delve' so much?
Reinforcement training rewarded a formal-helpful register, and words like 'delve' rode along — it became the most famous AI tell of 2023–24. Newer generations use it less (our dated corpora track the decline), but the installed base of old AI text still carries it everywhere.
If I use these words, does my writing look AI?
One word means nothing — people wrote 'tapestry' long before language models. The tell is density: several of these clustering in one passage, in any register, is what moves a score. Our scanner weighs the pattern, not the presence.
Should I stop using these words in my own writing?
Don't write scared. Cut them where they're filler ('in today's rapidly evolving landscape' says nothing), keep them where they're precise. Machine-sounding writing is a rhythm-and-cliché problem, not a banned-words list.
How is this list kept up to date?
We keep dated corpora of default-voice output from each major model generation, mine what changed, and hand-curate additions — words join only if they read machine-written in any register. When a model drops a word we demote it rather than delete it, because last year's AI text is still being pasted today.