AI writing has a recognizable vocabulary. Words like delve, tapestry, underscore, and leverage now appear far more often in machine-written text than in human writing, and enough readers have noticed that these words have become a folk signal for "this was written by ChatGPT." The pattern is real and measured: a 2024 study tracking 14 million scientific abstracts found a sharp, sudden rise in a specific set of style words right after ChatGPT launched. Below is the working list, grouped by type, with one important caveat up front: these are probabilistic tells, not proof, and treating them as proof is how human writers get wrongly accused.
Key facts
- A 2024 study (Kobak et al., Science Advances) analyzed over 14 million PubMed abstracts and found an abrupt post-ChatGPT surge in specific style words.
- A follow-up analysis found 103 of 135 candidate "AI words" rose at statistically significant rates after ChatGPT, with "delve" the single strongest marker.
- Wikipedia maintains a community-edited list, "Signs of AI writing," cataloguing the same vocabulary and punctuation tells.
- The signals drift: "delve" peaked and then faded through 2025 as models and prompts changed, so any fixed word list goes stale.
- These tells appear in ordinary human writing too, which is why detectors built on them produce false positives.
What words make writing sound like AI?
The most-cited offenders cluster into a few groups.
- Overused verbs and nouns: delve, leverage, foster, underscore, tapestry, realm, navigate, harness, pivotal, robust, intricate, nuanced, multifaceted, landscape ("the evolving landscape"), testament, beacon, cornerstone, embark, elevate, unlock, streamline.
- Hedging and filler phrases: "It is important to note that," "In today's fast-paced world," "When it comes to," "Plays a crucial role in," "A wide range of," "It is worth mentioning," "In the realm of," "Navigating the complexities of."
- Closing and transition tics: "In conclusion," "Ultimately," "Moreover," "Furthermore," "That being said," and the formulaic "Whether you are X or Y, there is something for everyone."
- Structural tells: the "not only X, but also Y" construction, groups of exactly three ("clear, concise, and compelling"), and rhetorical question-and-answer pairs.
Is the em dash really an AI tell?
The em dash has been nicknamed the "ChatGPT hyphen," and many readers now treat a post full of em dashes as a sign of AI. This belief is widespread enough that mainstream outlets and even OpenAI have addressed it. But it is folk wisdom, not a reliable test. Plenty of skilled human writers love the em dash, and some AI models use it sparingly. The em dash is worth knowing about because readers react to it, not because it proves anything. If your audience associates it with AI, that perception is the problem worth solving, regardless of who actually wrote the text.
Why these words signal AI
Language models choose the most probable next word at each step. That tendency pulls them toward common, safe, slightly formal vocabulary and steady sentence rhythm, which is exactly the cluster of words above. It is the same low-variation pattern that AI detectors measure. For the mechanism, see how AI detection works.
A warning: these are tells, not proof
This matters more than the list itself. The same vocabulary shows up in normal human writing, especially formal, academic, or non-native English writing. That overlap is precisely why AI detectors wrongly flag real people, sometimes more than 60% of the time for non-native writers. Do not use this list to accuse anyone, and do not panic if your own writing contains these words. For the full picture, see can AI detectors be wrong and our AI detector accuracy reference. A list of tells is a guide to sounding less generic, not a lie detector.
How to find and fix AI tells in your writing
If you want your writing to read less like a default AI draft, the fixes are concrete: cut the filler phrases, vary your sentence length, replace the generic verbs with specific ones, and add a detail only you would know. Metric37's free text tools include an em-dash remover for the punctuation tell, the free AI detector shows how your text scores, and the humanizer rewrites a draft into something that reads like a person wrote it, with a 0 to 100 human score so you can see what changed.