It’s 3 AM. You have a 2,000-word essay due tomorrow morning, and you just pasted your draft into an AI detector. The result? 78% likely AI. Your stomach drops.
You didn’t use AI to write the paper. But your professor read it anyway and asked if it sounds like your usual voice. Maybe it does sound a bit… stiff. Like you’re trying too hard to sound academic.
You’ve been told to “write formally.” So you reach for words you know professors like. You write “In this paper, I will delve into the multifaceted nature of this pivotal topic.” The sentence is grammatically correct. It also happens to be one of the most recognizable ChatGPT patterns on the internet.
Here’s what’s actually going on — and how to fix it.
The word “delve” has become the most famous ChatGPT giveaway on the internet. The numbers are staggering and they’re not anecdotal — they’ve been measured.
Data from Slop Or Not (slopornot.ai/words) shows the term appeared at 7.5 per 100,000 PubMed abstracts in 2022. By 2024, that number jumped to 357 per 100,000 — a 47.8× increase. The word didn’t exist at that frequency in academic writing before ChatGPT; the model made it viral. A single paragraph with one hit proves nothing, but three or four? That’s when reviewers start asking questions.
Other words with equally striking spikes include “showcase” (13.8×), “meticulously” (10.5×), and “underscore” (5.8×). These numbers run back to 2010, showing that humans have always used these words — the change is in frequency, not invention.
Reviewers have been “marinated in AI output for three years” and can instantly recognize a text that leans on several of these patterns at once. The moment a sentence like “this robust framework underscores a pivotal shift in the ever-evolving landscape” hits their eyes, their brain hits an AI alarm before they’ve even finished reading.
ChatGPT overuses certain words at rates that make the pattern instantly recognizable. Here’s the working list, grouped by what the words are doing in your text.
These words alone can spike an AI detector:
If your opening paragraph contains all three, the detector won’t need the rest of the paper.
These are ChatGPT’s structural backbone:
Words that sound impressive but read as formulaic:
How did “delve” become such a ChatGPT giveaway? The answer comes from a 2025 paper at the COLING conference (Juzek & Ward, https://aclanthology.org/2025.coling-main.426/) and reporting from The Guardian (Alex Hern, April 2024, https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt).
The COLING researchers tested the obvious explanations — training data, model architecture, algorithm design — and found none of them explained the pattern. What fit was the last stage of training: reinforcement learning from human feedback (RLHF).
During RLHF, human annotators rate the model’s outputs and tell it what sounds “good.” But the annotators aren’t working from Silicon Valley. The Guardian reported that AI companies send this feedback work to English-speaking workers in the Global South — Nigeria, Kenya, and similar regions — where the work is cheap to hire. In Nigerian business English, “delve” is a common and earnest word. It sounds polished. It sounds professional. The annotators rated it highly, and the model learned: “delve” = good writing.
The two explanations don’t compete. The annotators used “delve” because that’s what they were taught; the model learned that “delve” = professional; students and reviewers then noticed the pattern and turned it into a meme.
If you’re a non-native writer, the word might be exactly how you were taught to connect ideas in formal English. You didn’t copy ChatGPT — you copied a textbook. And that textbook probably used the word decades before ChatGPT existed.
This is the part most word lists skip entirely, and it matters enormously for the student audience.
Words like “delve,” “moreover,” “foster,” and “tapestry” are classic formal-register academic vocabulary taught in English textbooks across Eastern Europe, Asia, the Middle East, and Africa. IELTS and TOEFL preparation courses still coach linking words as a path to a higher score. The exact skills ESL students develop through years of English language study — formulaic transitions, uniform structure, safer word choice — are the same patterns AI detectors misinterpret as machine-generated text.
The evidence is not theoretical. A Stanford study (Liang et al., published in Patterns, 2023, https://doi.org/10.1016/j.patter.2023.100779) found that seven AI detectors falsely flagged 61% of TOEFL essays written by non-native English speakers. The detector was “near-perfect” on US eighth-grade essays but misclassified 61% of the TOEFL essays. The gap wasn’t a rounding error — it was a chasm.
Turnitin’s headline false positive rate is ~4% overall, but independent research shows ESL writers face rates 2-3× higher (~8-12%). In a 650-word essay, a native English speaker might see 1-2 flagged sentences; an ESL writer could see 3-6 flagged sentences from the same text length.
If you’re writing formally in English and your first language isn’t English, you should read this article carefully. The words you were taught to sound “academic” are the same words that make AI detectors flag your writing. This is sometimes called “flagxiety” — the low-grade dread of writing something that trips a detector even though every word is yours.
Here’s the complete list of chatgpt words to avoid in your academic writing. Use this as a quick reference — bookmark it or screenshot it for whenever you’re editing a draft:
| AI Word ❌ | Human Alternative ✅ | Example Sentence |
|---|---|---|
| delve | dig into, explore, look at | Let’s explore the connection between the two variables rather than using “delve into the relationship.” |
| tapestry | blend, mix, combination | A blend of cultural influences shaped the community’s identity, not “a rich tapestry of influences.” |
| testament | evidence, proof | The data provide clear evidence of the effect. |
| moreover | furthermore, additionally, also | The sample size was large; furthermore, the results were consistent across groups. |
| foster | encourage, develop, promote | The program encouraged students to participate actively. |
| pivotal | key, central, critical | The study provides a key insight into how students process feedback. |
| robust | thorough, strong, solid | The researchers conducted a thorough analysis of the results. |
| navigate | work through, explore, address | The study explores the challenges of implementing the new policy. |
| harness | use, take advantage of | The team used multiple sources to verify the findings. |
| nuanced | careful, detailed, subtle | The careful analysis revealed subtle patterns in student responses. |
| multifaceted | complex, varied, many-sided | The problem is complex and involves multiple factors working together. |
| seamless | smooth, easy, consistent | The transition between methods was smooth and well-organized. |
| leverage | use, take advantage of | We used the data from the previous study to design our experiment. |
| facilitate | help, enable, make easier | The software helped students organize their notes more efficiently. |
Use this table whenever you spot one of the flagged words. But stop here if you’ve ever swapped a word and thought, “I replaced it with something simpler — why does this still sound like AI?”
The answer is that the vocabulary swap is the visible symptom. The signal detectors and reviewers notice is structural — the rhythm, the shape, the information density. Swapping words without fixing the underlying structure is like repainting a car with a broken engine: it looks different, but it still won’t run.
Take a paragraph of AI output and manually replace every listed word. Read the result. It still sounds like AI — slightly cheaper AI, but unmistakably AI. This proves that vocabulary alone isn’t the signal.
Here is a simple test. Take a paragraph of pure AI output and manually replace every listed word. Read the result. It still sounds like AI — slightly cheaper AI, but unmistakably AI.
The words were never the core signal. They are the visible symptom of deeper patterns:
Uniform sentence rhythm. AI-generated text produces sentences of remarkably similar length and shape, each carrying roughly the same information load. Human writing “lurches.” Some sentences run long because the thought is complicated. Some are three words. AI output reads like a metronome; detection tools measure this evenness as “burstiness.”
Symmetric structure. AI loves triads (“clear, concise, and compelling”), balanced clauses (“not only X, but also Y”), and paragraphs that open with a topic sentence, deliver exactly three supporting points, and close with a mini-summary. Every paragraph. It’s the structure of a well-behaved school essay, executed with inhuman consistency.
Frictionless generality. AI text is smooth because it commits to nothing. It gestures at “various factors” and “numerous studies” without naming one. It hedges symmetrically (“while X has benefits, it also poses challenges”). A human writer who actually knows the subject cannot help leaking specifics — a date, a name, a number, an opinion that could be wrong.
No stakes. Nothing in the text costs the writer anything. There is no sentence where you feel the author risked being disagreed with.
Swap the vocabulary and all four properties survive untouched. This is why cargo-cult word avoidance fails: it imitates the surface of human writing without the mechanism. Worse, if you contort your text to dodge a word list — replacing natural words with awkward synonyms — you drift toward the “tortured phrases” territory that researchers have documented in academic publishing, where thesaurus-driven rewriting produced absurdities like “counterfeit consciousness” in place of “artificial intelligence.” Mechanical word substitution doesn’t make text more human. It makes it strange in a new way.
If the signal is rhythm and specificity, that is what you edit for. Here are five concrete techniques you can apply to any draft:
1. Vary sentence length on purpose. After a long sentence, write a short one. Let one paragraph carry a single two-line thought while the next runs eight lines. If you read your draft aloud and it sounds like a steady march, break the step.
2. Add details only you could add. Your data. Your example from last Tuesday. The specific thing your professor said. The number from your own experiment. Every concrete, checkable detail is a sentence no model would have produced.
3. Keep an opinion in the text. “This method is slower and I would not use it for anything over a thousand records” sounds human precisely because it is falsifiable. Symmetric hedging is machine music. You don’t need to be provocative — just take a stance.
4. Cut the ceremony, keep your register. Delete empty transitions (“In today’s fast-paced world, it is important to note that…”) but keep the formal register you learned. Don’t perform casualness that isn’t yours.
5. Watch the passive pile-up. Find places where agents disappeared. Restoring “we measured” instead of “measurements were conducted” does more for naturalness than any vocabulary swap.
And here’s the most important thing: the goal is “formal but human,” not “casual.” Academic writing already expects formality. AI words compound formality to an unnatural degree. You don’t need to rewrite your entire essay in a conversational tone to make it sound human — you need to vary the rhythm, add specifics, and let one sentence carry the weight of an argument.
You can use Grammarly to fix grammar and spelling. You should not use it to rewrite your sentences. Most universities make this distinction, and the line between them matters.
How silence works: University of Maryland uses a “silent brief” rule — AI is not allowed unless your syllabus says otherwise. University of Helsinki requires students to disclose use or non-use of AI in written work. Caltech’s Fall 2027 guidelines explicitly list “Translating an essay written in another language” as unethical use of AI.
Before you use any AI tool to edit your work, check your course syllabus and your institution’s policy. Read our guide on ChatGPT policy for a complete overview of responsible use, and check how to use Grammarly effectively without crossing the line into academic dishonesty.
Here’s a question most guides don’t answer: if AI detectors are unreliable (which the Stanford study proves), who is actually spotting AI writing?
The answer is people who work with AI daily. Russell et al. (2025) (https://arxiv.org/abs/2501.15654) tested 300 articles and found that reviewers who use ChatGPT for writing tasks detected AI text at 92.7% true positive rate with only a 4% false positive rate — outperforming even Pangram, the most accurate commercial detector.
Their primary cue was vocabulary: “crucial, vibrant, testament, delve, tapestry” showed up in 53.1% of their explanations. But they also noticed patterns that tools miss — quotes that all sound the same, overuse of generic names (“Emily,” “Sarah”), optimistic vague conclusions, and tonal inconsistency. These are higher-order signals — originality, voice, tonal coherence — that automatic detectors cannot reliably assess.
The takeaway: your professor is a much better detector than any online tool. That’s why fixing the structural patterns that make writing sound artificial matters more than dodging individual words.
This is one of those problems that feels overwhelming when you’re staring at a draft three hours before a deadline. You’ve been told to “write formally” and you’ve swapped “use” for “utilize” and “moreover” for “furthermore” and now your paper sounds like a robot wrote it — a textbook case of those chatgpt words to avoid that reviewers recognize instantly.
Here’s the fix: don’t try to fix everything at once. Pick one thing. Read your draft aloud — your mouth will stumble over formal passages that feel stilted. Rewrite one paragraph so it sounds like you’re explaining the idea to a peer. Add one detail only you could add — a number, a name, a specific example from your reading. The rest will follow.
If you want more help, check our paraphrasing techniques guide for practical strategies that go beyond word swapping, review our academic writing tone guide for discipline-specific expectations, and read about ChatGPT fake citations if you’re using AI tools for research help.
Still feeling unsure? Our writers have reviewed thousands of drafts and can tell you exactly what to change. See our guarantees for what’s included in every review.
No. “Delve” is the most famous, but the Slop Or Not data shows at least 35 words with significant spikes. Reviewers notice clusters — if you use three or more AI words in one paragraph, it’s worth a slower read.
Yes. The Stanford study proves that AI detectors misfire on formal English regardless of whether it was written by a human or an AI. Keep your formal register — it’s the right tone for academic writing — and focus on varying sentence length and adding specific details instead of chasing word lists.
Useful, but limited. Detectors can give you a heads-up, but they’re unreliable by design — especially for non-native writers. Better to read your draft aloud, check for structural patterns (uniform rhythm, no specifics, no stakes), and fix those directly.
Proofreading means catching spelling, grammar, and clarity errors. Co-authoring means asking an AI to rewrite sentences, alter your tone, or polish whole paragraphs. Most universities treat them as completely different things — one is acceptable, one is not.
Because the data is so clear. “Delve” jumped 47.8× in academic abstracts after ChatGPT arrived. The spike is measurable, not subjective. Other words like “showcase” (13.8×) and “meticulously” (10.5×) rose sharply too, but “delve” is the one reviewers recognize instantly because it became a meme on the internet.