AI models develop favorite words. "Delve" is the famous one, but the broader smell is the cluster: "robust," "leverage," "showcase," "pivotal," "tapestry," and newer model-era tics like "load-bearing."
The Pattern
Vocabulary tics are words and phrases a model reaches for more often than the surrounding human register would. One "robust" is normal. Three "robust" claims, two "pivotal" moments, and a "delve" in the same page is a fingerprint.
"Delve" is the best-studied case. FSU researchers Tom Juzek and Jeremy Ward traced the spike across millions of PubMed abstracts and tied it to ChatGPT's November 2022 release. Kobak et al. went wider: 15 million PubMed abstracts, 379 excess style words, and a measurable distribution shift in scientific writing after ChatGPT arrived.
The words change by model and era. GPT-4 had "delve," "tapestry," and "meticulous." GPT-4o leaned into "highlighting," "showcasing," and "emphasizing." Claude has its own visible habits in editing and coding prose; "load-bearing" is one of the newer field-observed examples. It is useful because it names a real concept, then becomes suspicious when every assumption, paragraph, and transition is suddenly load-bearing.
Examples
Tic Families
The classic cluster includes "delve," "intricate," "meticulous," "pivotal," "showcase," "underscore," "realm," "harness," "leverage," "facilitate," and "foster." These are not banned words. They become a smell when they arrive together and raise the register above the actual subject.
"Tapestry," "landscape," "realm," "paradigm," "ecosystem," "underpinnings," and "camaraderie" turn concrete claims into high-gloss generalities. The PNAS/arXiv style study found extreme overuse for some of these, including "tapestry" and "camaraderie."
"Robust," "comprehensive," "multifaceted," "nuanced," "pivotal," "intricate," "meticulous," "groundbreaking," and "transformative" all promise importance before the writing earns it.
"Furthermore," "moreover," "additionally," "consequently," "notably," and "it is worth noting" often mark a paragraph that is being procedurally assembled rather than argued.
Wikipedia editors call out the model habit of avoiding plain "is." Generated prose often says a thing "serves as," "acts as," "stands as," or "functions as" something else. The sentence feels more formal without getting more precise.
Some tics are not yet quantified in papers but are obvious in repeated use. "Load-bearing" is in this bucket: a useful engineering and criticism term that starts to read as model-shaped when it spreads from one precise claim to every sentence doing structural work. Treat it as a density signal, especially in Claude-assisted prose.
The Research
Juzek and Ward (December 2024) analyzed millions of PubMed abstracts and found 21 "focal words" that spiked sharply after ChatGPT's November 2022 release. "Delve" had the most dramatic increase. The spike doesn't just correlate with AI use -- it correlates with who trained the AI. Nigerian English uses "delve" at far higher rates than American or British English, and OpenAI's annotation workforce was heavily Nigerian.
Kobak et al. (Science Advances, 2025) went bigger -- 15 million PubMed abstracts, 379 "excess style words." By 2024, at least 13.5% of all PubMed abstracts showed signs of AI processing. The contaminated vocabulary goes well beyond "delve": "underscores," "aligns," "realm," "showcasing," "facilitates."
Wikipedia's "Signs of AI writing" guide tracks vocabulary by model era. The list changes because vendors patch obvious tells and new defaults replace old ones. That is why a static ban list fails: the better move is watching for clusters that are too polished, too repeated, and too mismatched to the author's normal voice.
The cross-genre style work on LLM prose found that instruction-tuned models diverge from human writing in predictable ways, including elevated rates for words such as "tapestry," "camaraderie," and other prestige-register terms. The finding matters because it shows vocabulary tics are not just blog folklore. They are measurable distribution shifts.
RLHF works by having human annotators rate model outputs, then optimizing for higher-rated text. If your annotators share a linguistic background, their dialect preferences get amplified into the model's voice. The model has no idea it's absorbing Nigerian English conventions. It just knows "delve" gets rewarded.
Caught in the Wild
Thousands of published scientific abstracts now contain "delve" and its cohort. Not isolated incidents -- the FSU paper documented field-wide vocabulary distribution shifts starting in Q1 2023, visible across entire journals.
Kobak et al. study →An Elsevier paper in Surfaces and Interfaces shipped with the literal ChatGPT response prefix as its opening line. Unedited AI text, straight through peer review. Still published. Still unretracted.
Technology Networks →Wikipedia editors watched "delve" colonize thousands of new and edited articles. The WikiProject AI Cleanup team, founded in late 2023, now tracks vocabulary tells as a front-line detection signal. Their findings fed directly into Wikipedia's March 2026 ban on AI-generated article content, which passed 44-2.
TechCrunch →Sources