AI Ghostwriting Is Creeping Into Science—Is That a Bad Thing?

Which words give AI away? A new study of more than 15 million biomedical abstracts on PubMed found that at least 13.5% of scientific papers published in 2024 show signs of AI-assisted writing tools, most notably OpenAI’s ChatGPT.

The study by researchers from Northwestern University and the Hertie Institute for AI in Brain Health at the University of Tübingen found a sharp rise in 2024 in word patterns associated with AI-generated writing. These included both uncommon terms—such as “delves,” “underscores,” and “showcasing”—as well as more familiar words like “potential,” “findings,” and “crucial.”

To measure this change, researchers compared word frequencies in 2024 against baseline data from 2021 and 2022. They ultimately identified 454 words frequently overused by AI models, including “encapsulates,” “noteworthy,” “underscore,” “scrutinizing,” and “seamless.”

However, experts explained to Decrypt that word frequency alone isn’t sufficient evidence of AI use.

“Language changes over time,” said Stuart Geiger, assistant professor of communication at UC San Diego. “‘Delve’ has skyrocketed, and this word is now in the vocabulary of society, partly because of ChatGPT.”

Geiger emphasized that detecting AI in writing isn’t just a technical challenge; it’s also ethical.

“The only way to reasonably detect LLM use is if you’re there, surveilling the writing process,” he said. “That comes at a high cost, logistically, morally, and technically.”

However, Stuart warned against jumping to conclusions based on surface-level clues without knowing the full context.

“It could be they’ve just seen a bunch of ChatGPT-generated writing and now think that’s what good writing looks like,” he said. “That’s the whole issue that we in academia are struggling with, especially when we can’t just put students in seats and make sure that it’s just pen and paper.”

As AI-generated text becomes more common, educators have turned to tools that claim to detect it; however, the quality of these tools varies.

In October 2024, Decrypt tested leading AI detection tools—including Grammarly, Quillbot, GPTZero, and ZeroGPT. Results varied wildly: ZeroGPT claimed that the U.S. Declaration of Independence was 97.93% AI-generated, while GPTZero gave it just 10%.

“There’s a lot of snake oil being sold,” Geiger said.

According to Geiger, concerns about AI writing tools echo past debates over spell check, Wikipedia, and CliffsNotes, and reflect deeper questions about the purpose of writing, authorship, and trust.

“People are concerned that when you had to write the words yourself, you had to think about them,” he said. “That’s what people react so strongly to when they see something that feels suspect.”

Rice University Professor of Business Kathleen Perley argued that while AI writing often shows patterns, such as repeated structures or overused words like “delve,” what matters most is whether it helps researchers without compromising quality. That’s especially true, she said, for non-native English speakers or people facing other challenges.

“If AI helps researchers overcome challenges like language barriers or learning disabilities, and doesn’t compromise the originality or quality of their work, then I don’t see a problem with it,” she told Decrypt. “I think it could be an overall benefit because it’s allowing people who have had different backgrounds, ideas, exposures, to participate in something that might have been an obstacle because of a lack of formal writing skills.”

The AI advisor to the deans at Rice Business, Perley noted another dilemma is the inclination of people to change the way they write out of fear of being accused of using AI, adding that she has become more conscious of certain words that might be flagged as potentially AI-generated.

While some criticize this style for lacking personality, Perley sees AI-assisted writing as a tool that can democratize participation in formal research.

“Sure, we might get more ‘delves’ and em dashes,” she said. “But if AI helps people from different backgrounds share important research, I don’t care how polished it sounds—it’s worth it.”