AI-generated content is starting to show in scientific writing as researchers investigate the trends of AI use in manuscripts. Analyzing the use of 15 specific verbs in titles can sometimes be enough to predict AI-generated scientific content.
Our over-the-top embrace of AI-assisted content creation since OpenAI’s ChatGPT launch in 2022 might be coming back to haunt us. That innocent manuscript title, “Navigating the [enter your topic],” whipped up in seconds to meet an approaching deadline, now contains ugly and evident AI traces. We may no longer be immune to scrutiny; editors and readers might now see we took the lazy way out.
If that weren’t embarrassing enough, an AI-generated title even hints we’ve extended the same laziness to the rest of the content. That may be a stretch, but it brings us to a study by Dr. Ruben Comas-Forgas, a professor of Applied Pedagogy and Educational Psychology, and his colleagues. They examined how generative AI influences academic writing by analyzing manuscript titles and found a significant increase in the use of 15 specific verbs in titles published between 2023 and 2024.
Apart from revealing the 15 verbs we should avoid to prevent embarrassment, this study indirectly teaches us about academic writing, transparency, and the inclusion of scientists into academia.
The 15 verbs raising suspicion of AI-generated titles
Comas-Forgas first noticed an odd trend shift when reviewing manuscripts in his fields, education and information sciences. Specific words began appearing far more frequently in titles he received by 2023 than in earlier years. Since this increase coincided with the widespread adoption of ChatGPT, he wanted to know if AI assistance could explain the trend.
The researcher first compiled a short list of verbs he suspected were overused in AI-generated titles based on his experience reviewing manuscripts. He then expanded the list by feeding 40 published abstracts into ChatGPT-4, asking it to generate five titles for each. The 200 titles generated by ChatGPT contained additional “suspicious” verbs.
He and his colleagues then cross-referenced these verbs with article titles in three major academic databases (SCOPUS, Web of Knowledge, and DOAJ), tracking their frequency over time, particularly before and after 2023. They identified 15 verbs whose use had significantly increased, suggesting a link to AI-assisted writing.
The results are straightforward. The study estimated that several titles (63,780 over two years) were potentially AI-generated or assisted. Comas-Forgas and colleagues identified four verbs in published titles with an annual growth rate of more than 100%: revolutionizing (381%), unleashing (208%), unveiling (157%), and unlocking (137%).
High (above 100% annual growth)
- Revolutionizing: 381%
- Unleashing: 208%
- Unveiling: 157%
- Unlocking: 137%
Moderate (50–100% annual growth)
- Unraveling: 89%
- Navigating: 74%
- Uncovering: 71%
- Advancing: 66%
- Harnessing: 63%
- Enhancing: 63%
- Pioneering: 54%
Low (below 50% annual growth)
- Deciphering: 44%
- Exploring: 42%
- Leveraging: 29%
- Bridging: 24%
They also pinpointed the fields and geographic locations where scientists prominently used AI for title generation. Scientists from multidisciplinary studies, sciences and engineering, and health sciences had the most notable increase in potentially AI-generated or AI-assisted titles. And perhaps less surprisingly, scientific authors from non-English speaking countries had adopted AI use for title generation much more (about 145%) than authors from English-speaking countries (about 60%) when comparing 2021-2022 and 2023-2024 data.
AI ruins academic writing
We might intuitively think that AI use threatens scientific writing and emphatically support any initiatives limiting it.
The ethical and transparency issues related to undisclosed AI use are real. Comas-Forgas and colleagues highlight these issues accurately, “Transparency in disclosing the use of AI in the writing process will be key in maintaining academic integrity and trust in scholarly communication.” Trust in academia relies on transparency, so disclosing any AI use—even when used responsibly—can make a significant difference in the perception of science.
Another issue in scientific dissemination concerns AI hallucinations, where the tool introduces biases and factual errors. Although internal and peer reviews can catch and correct these before publication, subtle inaccuracies can go unnoticed.
But, the study’s worries about the degradation of the academic voice require more in-depth analysis and discussions. They are more complicated and nuanced than we may think.
I’ve previously argued that AI can function as a cast for our writing and critical thinking and that we should remain vigilant against its overuse to protect our creativity. AI may exponentially contribute to a uniform and dull content landscape through a feedback loop that eventually leaves little space for originality.
What is a unique academic voice, anyway?
Comas-Forgas and colleagues express similar concerns, “The uniform increase in certain verbs across different fields suggests that [generative AI] might be nudging academic titles towards a standardized format, potentially reducing the individuality and innovation that are hallmarks of scholarly knowledge.” They couple these sentiments with worries about losing what they refer to as an “individual academic voice.”
Although we may want to maintain high academic standards and feel these fears are straightforward and self-evident, they might not be. AI’s negative effect on the academic voice and individuality is an interesting and debatable perspective. Let’s tackle its ambiguity.
Academic writing already entails standardized formats that depend on the fields. Academic writing is formal, unbiased, explicit, focused, structured, well-sourced, and consistent. These features prevent academic literature from becoming sensational or exaggerated; they maintain the arguments as objective as possible.
But let’s assume it’s true that academic writing expresses the unique voice of a researcher, which vanishes with AI use. Comas-Forgas and colleagues imply that standardizing the writing threatens an “individual academic voice,” but this raises a question: Why is an “individual academic voice” necessary for strong academic writing?
We can ask ourselves how uniform writing negatively affects science and scientific dissemination. If it does, then large swaths of academic writing are already in trouble if you ask me.
But this all assumes that AI language hampers academic writing.
AI for a more inclusive academia
The negative effect of generative AI on academic or scientific writing is not a language issue. If anything, I’d claim that AI can make publishing fairer. We saw that scientists living in non-English speaking countries contributed to the most pronounced annual increase in suspected AI-generated titles. These scientists have long had a disadvantage in an English-dominated work field. We could argue that generative AI levels the playing field for non-native English speakers.
I recently critiqued replacing the current publishing with a club model where scientists from a specific club (field) form a selective environment to disseminate scientific knowledge. I find this model problematic as a replacement because it tends to limit multidisciplinary and interdisciplinary research.
Academic writing is challenging and, although somewhat standardized, field-specific. Even native English speakers have issues writing or reading academic content, especially if they’re new to a field or the articles contain unnecessarily dry and unengaging writing. It may not be a coincidence that scientists from non-English countries or multidisciplinary fields use AI. The technology makes field-specific language and style more accessible, increasing these scientists’ chances to succeed.
Connecting AI language and transparency
Let me propose a solution that many have suggested before me. What if we accept AI-assisted scientific manuscripts to help with language—let it become dull (isn’t it already?)—but on one condition: scientists must disclose its use. Disclosing AI use will warn peer reviewers to scan for AI-specific biases or errors.
By accepting the responsible use of AI, we destigmatize scientists’ attempts to join the club and invite more scientists to succeed in the scientific publishing game.
Overly AI-specific language will be selected out by reviewers in any case, and, as in the case of science communication or critical thinking, the uniquely crafted AI-free content will stick out from the dull, standardized writing. Or am I being too naïve?
In the meantime, we can do ourselves a favor and avoid the 15 verbs above. Print them out and hang them above your screen; they reek of AI.
