Detecting AI-generated text has long been a challenge for researchers and developers. With the rapid advancement of large language models (LLMs), such as Google’s Gemini Advanced and OpenAI’s GPT-4o, the ability to produce human-like text has become increasingly sophisticated.
However, a new study from researchers at the University of Tübingen and Northwestern University offers a breakthrough in identifying AI-crafted content.
By focusing on the sudden surge in specific vocabulary in scientific writing, they have developed a method to detect the use of LLMs with surprising accuracy. This technique, inspired by pandemic studies that measured excess deaths, reveals how changes in word usage can signal the presence of AI-generated text.
![Avoid these words at all cost if you don't want to get caught using AI 1 Here are common words in AI-generated content](https://techbriefly.com/wp-content/uploads/2024/07/Here-are-common-words-in-AI-generated-content-_1.jpg)
What are the words that give AI content away?
To measure these changes, the team scrutinized the frequency of each word annually. By comparing the expected word frequency, based on pre-2023 trends, to actual usage in 2023 and 2024, they identified a dramatic increase in certain terms. For example, the word “delves” appeared 25 times more frequently in 2024 abstracts than anticipated. Similarly, “showcasing” and “underscores” saw a ninefold increase in usage.
Here are the most used words in AI-generated text with their corresponding rates of increase in usage:
- Delves – 25 times increase
- Showcasing – 9 times increase
- Underscores – 9 times increase
- Potential – 4.1 percentage points increase
- Findings – 2.7 percentage points increase
- Crucial – 2.6 percentage points increase
- Across – significant increase (exact rate not specified)
- Additionally – significant increase (exact rate not specified)
- Comprehensive – significant increase (exact rate not specified)
- Enhancing – significant increase (exact rate not specified)
- Exhibited – significant increase (exact rate not specified)
- Insights – significant increase (exact rate not specified)
- Notably – significant increase (exact rate not specified)
- Particularly – significant increase (exact rate not specified)
- Within – significant increase (exact rate not specified)
These words have become telltale signs of AI involvement, showing up far more frequently than expected. While language evolves naturally, such abrupt changes are unusual and often tied to major global events.
In this case, the widespread use of LLMs has led to a noticeable shift in the vocabulary of scientific literature.
Inspiration from pandemic analysis
The researchers’ approach draws heavily from techniques used during the COVID-19 pandemic. Just as excess deaths were calculated by comparing observed fatalities to historical data, this study compares current word usage against historical trends to identify anomalies. They analyzed over 14 million scientific abstracts published on PubMed from 2010 to 2024, identifying a significant uptick in certain words starting in late 2022, coinciding with the broader adoption of LLMs.
The researchers noted that the rise in specific words, termed “marker words,” is a clear indicator of LLM usage. This phenomenon differs from past vocabulary shifts linked to events like the COVID-19 pandemic, which saw an increase in noun-heavy language.
![Avoid these words at all cost if you don't want to get caught using AI 2 Here are common words in AI-generated content](https://techbriefly.com/wp-content/uploads/2024/07/Here-are-common-words-in-AI-generated-content-.jpg)
In contrast, the post-LLM period has seen a spike in verbs, adjectives, and adverbs. This shift highlights how AI-generated text subtly changes the texture and style of writing.
By identifying these marker words, the researchers estimate that at least 10% of scientific abstracts in 2024 were either generated or significantly assisted by LLMs. This estimate is likely conservative, as not all AI-assisted texts will contain these specific markers. Nonetheless, the presence of these words provides a reliable metric for detecting AI influence in academic writing.
Geographical trends in LLM usage
The study also uncovered geographical variations in the adoption of LLMs. Countries like China, South Korea, and Taiwan showed a higher frequency of marker words in scientific papers, indicating that LLMs are particularly valuable for non-native English speakers. These tools help refine and enhance their writing, making it more polished and publication-ready.
Conversely, native English speakers may be more skilled at recognizing and eliminating these markers, thereby concealing their use of AI. This difference suggests that while LLMs are widely used across the globe, their impact is more pronounced in regions where English is not the primary language.
Featured image credit: Freepik