How AI is Secretly Rewriting Scientific Papers: Tracking the Linguistic Footprints of LLMs in Biomedical Literature

If you have read a scientific paper recently, you might have noticed a strange shift in how they are written. It is not that the science has suddenly changed, but rather the vocabulary. A groundbreaking study highlighted by Science | AAAS reveals that we can actually track the massive influx of large language models (LLMs) in academic publishing simply by looking at the sudden explosion of specific, fancy words. Let us look at how researchers are using "excess vocabulary" to map the footprint of AI in top-tier biomedical journals, and what this means for the future of scientific integrity.

The Tell-Tale Vocabulary of AI in Science
Tracking the Sudden Surge of "Delve" and "Meticulous"
My Own Experience with AI Word Bloat
What Excess Vocabulary Means for Scientific Integrity
How to Collaborate with LLMs Without Losing Your Voice

The Tell-Tale Vocabulary of AI in Science

For decades, academic writing has had its own distinct, often dry, style. But right around late 2022, something shifted. Suddenly, biomedical papers became incredibly fond of words like "delve," "testament," "meticulous," "pivotal," and "showcase." This was not a organic cultural shift among scientists worldwide. Instead, it was the direct result of researchers feeding their drafts into tools like ChatGPT to clean up their English, structure their abstracts, or write entire sections from scratch.

When researchers analyzed databases of millions of biomedical papers published before and after the generative AI boom, they found a statistical anomaly. The frequency of certain transition words and adjectives did not just grow; it skyrocketed exponentially. These words act like a digital fingerprint, showing exactly where an LLM was used to polish, edit, or generate text. Because these AI models are trained to be polite, thorough, and slightly dramatic, they default to a very specific set of vocabulary that most human writers rarely use so densely.

Pro-Tip: If your draft is filled with words like "furthermore," "intricate," and "underscores," you are probably letting the LLM take too much control over your natural voice.

Tracking the Sudden Surge of "Delve" and "Meticulous"

The statistical tracking of these "excess words" is actually quite fascinating. Scientists compared the word choices in post-2022 papers with a baseline of papers written over the previous two decades. The word "delving", for example, saw an increase of over thousands of percent in peer-reviewed biomedical literature. Under normal circumstances, vocabulary usage in a highly specialized field changes incredibly slowly. A sudden spike like this is the linguistic equivalent of a volcanic eruption.

Why do LLMs love these words so much? It comes down to how they are trained. Through Reinforcement Learning from Human Feedback (RLHF), AI models are rewarded for sounding helpful, comprehensive, and authoritative. This biases their internal pathways toward structured, slightly flowery language. They prefer transition words that sound highly professional but are actually just filler. When a scientist asks an AI to "make this paragraph sound more professional," the model simply injects these high-frequency AI favorite words, leaving behind a clear trail for detectors and style-conscious peer reviewers.

My Own Experience with AI Word Bloat

Honestly, I have tried this myself when editing technical reports and drafts. When I first started using Claude and ChatGPT to help clean up my writing, I thought the outputs looked incredibly polished. But after a few weeks, a sense of repetition set in. I realized every single summary I generated started with some variation of "This study delves into..." or ended with "This serves as a testament to..." It became incredibly annoying. I found myself spending more time editing out the "AI-isms" than I would have spent writing the transition sentences myself. It taught me that while LLMs are fantastic for brainstorming and structuring, their default vocabulary settings are incredibly generic and easy to spot from a mile away.

What Excess Vocabulary Means for Scientific Integrity

On one hand, using an LLM to polish a paper is a massive win for accessibility. Non-native English-speaking scientists face massive hurdles when trying to publish in prestigious, English-dominated journals. AI tools level the playing field, allowing brilliant minds from all over the world to present their data clearly without being rejected purely based on grammar or syntax. That is a genuinely positive shift for global science.

On the other hand, the sheer volume of excess vocabulary suggests that many researchers are not just editing; they are letting AI write large portions of their papers. If a researcher is too rushed to write their own discussion section, did they also rush the data analysis? This is where peer reviewers are starting to get nervous. The concern is that this linguistic homogenization is a symptom of a deeper issue where papers are being generated with minimal human oversight, leading to the potential spread of hallucinated data or poorly peer-reviewed conclusions.

"The issue isn't the AI assisting with the language; the issue is when the language replaces the critical, human thought process required in scientific research."

How to Collaborate with LLMs Without Losing Your Voice

If you want to use LLMs to help with your academic or professional writing without leaving a massive, obvious AI footprint, you need to change how you prompt them. Instead of giving a loose prompt like "improve this text," you should actively set boundaries. Tell the model to avoid its favorite crutch words. You can literally prompt it by saying, "Do not use the words delve, testament, meticulous, pivotal, or furthermore." This forces the AI to use simpler, more direct vocabulary.

Another great approach is to use the AI purely as an editor, not a writer. Ask it to find grammatical errors or structural logical leaps in your writing, rather than asking it to generate new paragraphs. Keep your personal voice intact. After all, science is driven by human curiosity, and our publications should sound like they were written by real people, not a sterilized, over-polished algorithm.

Frequently Asked Questions

Q: Why do AI writing detectors look for specific words instead of just analyzing grammar?

A: LLMs generate grammatically perfect text, so grammar alone cannot reveal AI usage. Instead, detectors and researchers look for statistical anomalies in word choice. When words like "delve" or "showcase" suddenly appear at ten times their historical rate in a specific field, it points directly to algorithmic generation.

Q: Is it unethical for scientists to use LLMs to write peer-reviewed papers?

A: It depends on how it is used. Using AI to improve readability, fix grammar, or translate text is generally seen as acceptable and even beneficial. However, using AI to generate hypotheses, interpret data, or write discussions without thorough human verification is highly controversial and often violates journal guidelines.

Q: How can I make my AI-assisted writing sound more human?

A: The easiest way is to edit out common AI transition words and simplify your sentence structures. Keep your prompts highly specific, explicitly tell the AI to use a conversational or direct tone, and always do a final pass yourself to put your unique voice back into the text.