The AI Words List: 120+ Phrases ChatGPT Overuses

1Why ChatGPT Overuses These Specific Words

This is not folklore. It is measured.

In 2024, a team led by Dmitry Kobak analyzed 14.2 million PubMed abstracts from 2010 onward. Certain "excess" words spiked so hard after ChatGPT's release that the authors estimate at least 13.5% of 2024 abstracts were processed with an LLM, reaching 40% in some subcorpora (Kobak et al., 2024). The most lopsided jump belonged to delves, which appeared at roughly 25 times its expected rate. Showcasing and underscores followed at about 9x each.

Zoom in on one word in a different database and the curve is just as steep. In the Dimensions corpus, delve went from 0.46% of abstracts in 2022 to 4.3% in 2024, a near-9x rise (Kobak et al., 2024). This is not a medicine problem. Across 25 leading economics journals, the frequency gap between LLM-flavored terms and normal economics language more than doubled in one year, from 2.85 points in 2023 to 6.67 points in 2024 (Feyzollahi & Rafizadeh, Economics Letters, 2025).

So why these words? The leading explanation is the training itself. Florida State University researchers identified 21 "focal words" that spiked with no natural cause, then reproduced the effect in the lab: models fine-tuned on human preference data grew fonder of these words than the base models did. As co-author Tom Juzek put it, "the enhanced model was less surprised by abstracts containing the buzzwords than the base model that had not been trained on human preference data" (FSU News, 2025). In plain terms, the people who rated model outputs during training appear to have quietly rewarded this register, and the model learned to write like an over-eager grant application. One honest caveat: the same team found delve behaves oddly compared with the other 20 words, so human feedback is a source, not the whole story.

That is the mechanism. Below is the vocabulary it produces. None of these words is banned, and plenty are useful in the right sentence. They are overrepresented, and when they pile up, a reader (or a detector) starts to smell the machine.

The tell is never one word. It is a room full of words all wearing the same suit.

2Tier 1: Kill on Sight (28 words)

These almost never show up in unforced human writing. When you see one, look for the others; they travel in packs. Each is flagged at full weight in our engine.

3Tier 2: Suspicious in Clusters (44 words)

Any one of these is fine. Three or four in a paragraph, especially the adjectives, is the sound of a model reaching for shine it hasn't earned. This is where most AI buzzwords live: the promotional register Wikipedia's editors formally flag as a sign of machine writing, right down to "boasts a" and "vibrant" (Wikipedia: Signs of AI writing).

4Tier 3: Light Signals (19 words)

These are ordinary words. You use them, I use them. They only count when they show up too often and too close together, especially the transition words, which AI sprinkles like seasoning. Do not purge these from your writing. Just notice when a paragraph has four of them.

5The Cliché Phrases: Where AI Really Gives Itself Away

Single words are the easy part. The stronger tells are phrases: fixed strings a model reaches for because they were rewarded a million times in training. They are worse than any single word because they drag a whole rhythm along with them. If you write for a living, memorize this section.

6AI Buzzwords: The Adjective-Plus-Noun Glue

One more layer earns its own callout, because it's what people mean when they search "ai buzzwords." The strongest slop signal often isn't a lone word. It's a collocation: two words that clip together and mean almost nothing. Our detector scores these as phrases, not as their component words, precisely because the glue is the tell.

7The Transition-Word Habit

The last pattern is structural rather than lexical. AI doesn't just overuse transition words; it uses them to start paragraphs, over and over: However. Furthermore. Additionally. Moreover. Consequently. Wikipedia flags "Additionally" as a signature paragraph-opener in earlier model output. One or two is fine. When four of five paragraphs open with a formal connective, the writing has a metronome where its argument should be. The fix is boring and effective: start some paragraphs with the actual subject.

8How to Use This List Without Becoming Paranoid

A blacklist is the wrong takeaway. If you delete every word above from your vocabulary, you'll write stiff, hollow prose, which is its own kind of slop. Here's the sane workflow:

The fastest way to see your own clusters is to let something count them for you. Our free, in-browser slop checker highlights every flagged word and phrase in your text and tells you why each one fired (Tier 1, Tier 2, cliché phrase, transition tic), so you fix the density instead of guessing. Nothing you paste is uploaded; the whole engine runs locally. And if you want the categories these signals fall into, the named families of AI slop from Fake Authority to Pseudo-Insight, read our Slop Taxonomy. That taxonomy is the framework this word list is built on.

Two companion reads while you're here. If you're worried your own writing has drifted into this register, How to Write Content That's NOT Slop is the practical fix. And for the full field guide to spotting AI across images and video too, start with How to Spot AI Slop.

Frequently Asked Questions