Recommended Free Tools
Yes, some words appear unusually often in particular bodies of writing associated with large language models (LLMs), especially scientific abstracts. Studies have highlighted words such as delve, intricate, and underscore. But a conspicuous word is not an “AI fingerprint”: frequency depends on the genre, date, editing, and comparison corpus, and one word cannot establish who wrote a passage.
What “AI slop” means in this evidence
“Slop” describes writing judged careless or low quality. The studies discussed here do not measure quality directly. They measure lexical frequency, excess vocabulary, or vocabulary diversity in defined collections of text. Their results support a narrower claim: some LLM-associated writing contains certain words more often than comparable earlier or human-authored writing.
As an Amazon Associate I earn from qualifying purchases.
That distinction matters. A person can use delve naturally, and an AI-assisted draft can be heavily rewritten. Conversely, a polished passage may retain statistical traces of model output. Word choice is evidence about a corpus, not proof about an individual author.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsWhich words have researchers identified?
Tom S. Juzek and Zina B. Ward’s COLING 2025 paper, “Why Does ChatGPT ‘Delve’ So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models,” identifies 21 focal words whose rising use in scientific abstracts is likely related to LLM usage. Its abstract names delve, intricate, and underscore as examples.
#1 Best Overall
The authors write: “Application of our method yields 21 focal words whose increased occurrence in scientific abstracts is likely the result of LLM usage.” They do not present a permanent blacklist. The relevant baseline is the surrounding corpus and period, not whether a word sounds unusually formal.
What the main studies measured
| Study | Corpus and period | What it found | What it cannot establish |
|---|---|---|---|
| Juzek and Ward, COLING 2025 | Scientific abstracts; changes over time | Twenty-one focal words became disproportionately common; examples include delve, intricate, and underscore. | It does not prove that a particular sentence was generated by AI or identify a single cause with certainty. |
| Geng and Trotta, Findings of ACL 2025 | arXiv paper abstracts; time-varying analysis | Delve and several publicized ChatGPT-associated words fell after early 2024, while significant continued to rise. | The pattern does not yield a universal detector; authors may select, edit, or adapt model output. |
| Scientific Reports study, 2024 | Application materials classified as AI-generated, AI-revised, or human-authored | The indexed abstract reports that AI-generated documents used a smaller vocabulary and repeated favored words. | The article’s full methods and precise figures were not available in the indexed material, so exact numerical claims should not be inferred. |
| PubMed-indexed biomedical study, 2025 | More than 15 million biomedical abstracts from 2010–2024 | Its excess-word analysis estimated that at least 13.5% of 2024 abstracts were processed with LLMs. | This is a method-dependent estimate for that corpus, not a count of disclosed AI use or a prevalence rate for all writing. |
Why the “AI words” list keeps changing
Mingmeng Geng and Roberto Trotta describe this as a human–LLM coevolution. After words such as delve were publicly identified in early 2024, their frequency in arXiv abstracts dropped. The authors state that this happened “starting soon after they were pointed out in early 2024.” Meanwhile, significant continued to increase.
This is consistent with several ordinary behaviors: writers may revise a draft, publishers may change style guidance, and users may prompt models to avoid recognizable phrases. A detector trained on yesterday’s list can therefore lose accuracy when people adapt. The finding also explains why a word can be statistically associated with model output in one period yet be unhelpful later.
Do LLMs write like humans?
They can produce grammatical and rhetorically plausible prose, but research finds measurable differences in style distributions. A 2025 study titled “Do LLMs write like humans? Variation in grammatical and rhetorical styles” provides broad qualitative corroboration that LLM and human writing can vary in grammatical and rhetorical patterns. That does not turn stylistic variation into a reliable authorship test.
Published text also mixes sources. An abstract may be generated, translated, lightly edited, substantially rewritten, or assembled from several authors’ work. Observed vocabulary can reflect any combination of those steps.
What may cause lexical overrepresentation?
Juzek and Ward tested possible explanations rather than assuming that a model simply “likes” certain words. Their analysis found no evidence that model architecture, algorithm choices, or training data explained the pattern. Model testing was consistent with reinforcement learning from human feedback (RLHF) contributing, but the paper treats that result as suggestive, not conclusive. Limited transparency around model development leaves the causal question unresolved.
Rank #4
That uncertainty is important: a frequency shift can be real without its mechanism being settled.
Can one word prove that text was AI-written?
No. A word becomes informative only as part of a documented comparison: a defined genre, a time window, a representative baseline, and a method that examines many features. Even then, the result is probabilistic and corpus-bound.
Best Value
- Improve and refine your student's sentence and paragraph skills
- Lessons and activities progress from writing sentences to writing paragraphs
- There are complete teacher instructions and over 70 reproducible models and student writing forms
- Grades 4-6
- 136 pages
- Do not treat delve, intricate, underscore, or significant as forbidden words. They are ordinary English words whose usefulness as signals depends on context and frequency.
- Do not infer authorship from one sentence. Human writers can produce the same wording independently.
- Separate generation from assistance. A human-edited passage may retain some model preferences without being fully generated.
- Check the corpus. A result from biomedical or arXiv abstracts should not be generalized automatically to news, fiction, student essays, or workplace email.
- Account for time. Public attention can change the very frequencies a detector relies on.
How to evaluate a suspicious passage responsibly
- Read for meaning first. Look for unsupported claims, circular explanations, fabricated citations, or generic transitions; these are quality issues, not proof of AI use.
- Look for a pattern, not a token. Repeated vocabulary, unusually narrow word diversity, and formulaic rhetorical structure may justify closer review, but none is decisive alone.
- Compare like with like. Use writing from the same field, publication type, language variety, and period whenever a statistical comparison is possible.
- Ask about the writing process. Draft history, citations, notes, and revision records can provide stronger evidence of authorship than stylistic suspicion.
- Report uncertainty. Describe a passage as showing features associated with a studied corpus, not as definitively AI-generated.
What readers should take away
Research supports a measured conclusion: LLM-related writing can overrepresent particular words, and those patterns are detectable at corpus scale. The examples and their frequencies move over time as people notice and edit them. Because the studies cover specific genres and mixtures of human and machine assistance, lexical clues cannot authenticate an individual passage.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
