Did LLMs colonize scientific titles? Fifty years of titular colonicity
This study analyzes fifty years of scientific title data to demonstrate that while colon usage has been a longstanding convention that is rising in prevalence, large language models generate titles with significantly higher colon frequencies than human authors, a pattern that is prompt-sensitive and influences retrieval ranking performance.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Scientific papers are often judged by their titles, those brief phrases that must do the heavy lifting of telling a reader what a study is about before they even click a link. For decades, researchers have noticed a specific pattern in these titles: the use of a colon, that small punctuation mark consisting of two dots. A title might read "Climate Change: A New Perspective," using the colon to split a broad subject from a specific focus. This habit, known in academic circles as "titular colonicity," has been observed since the 1980s. It is a standard tool for organizing information, but recently, a new force has entered the room: large language models. These are artificial intelligence systems trained on vast amounts of text that can now write summaries, answer questions, and even draft scientific titles. As these tools become common in research, a natural question arises: are they changing the way we write? Specifically, are these machines overusing the colon, perhaps making it a signature of their output, and does this punctuation mark actually help or hurt when a computer tries to find a paper in a database?
To answer this, a team of researchers from Interactive Crafts in Poland conducted a massive investigation spanning fifty years of scientific history. They did not just guess; they looked at nearly ninety-five million journal articles published between 1976 and 2025, a collection drawn from a global database called OpenAlex. Their first task was to map the history of the colon. They found that its use has never been a straight line up or down. Instead, it has wavered over the decades, sitting at about 16 percent in the late 1970s, rising and falling, and then climbing again in recent years. By 2022, roughly 25 percent of English-language titles contained a colon. By 2024, that number had jumped to nearly 29 percent, and a preliminary look at 2025 suggested it might reach 34 percent. While this recent rise is noticeable, the researchers were careful to note that this trend started before artificial intelligence became a dominant tool for writing, meaning the increase is part of a longer, complex story of how scientists write, not necessarily a sudden shift caused by machines.
The team then turned their attention to the machines themselves. They took nearly 1,700 scientific abstracts—summaries of research papers—and asked three different open-source language models to generate new titles for them. They gave the models two types of instructions: one was neutral, simply asking for a title, and the other asked the models to follow the specific style conventions of the scientific field, such as biology or engineering. The results were striking. When the models wrote titles, they used colons far more often than human authors do in published papers. For the primary model tested, 41 percent of the neutral titles contained a colon, compared to only 27 percent of the actual published titles. When the models were told to follow field conventions, the rate of colon use jumped even higher, reaching 60 percent. This suggests that while humans use colons moderately, these artificial intelligence systems have a strong tendency to overuse them, especially when prompted to mimic specific academic styles. The researchers also found that the models were inconsistent; sometimes they would produce a title with a colon, and other times they would not, even when looking at the same abstract, indicating that this behavior is sensitive to how the question is asked.
Finally, the researchers wanted to know if this punctuation actually mattered when a computer tried to find a paper. They set up a controlled experiment where they took titles that had a single colon and replaced that colon with a simple space, leaving every other word exactly the same. They then asked a search system to find the original paper using the modified title. Surprisingly, removing the colon made the search slightly more effective. The system found the correct paper more often and ranked it higher in the list of results when the colon was gone. This does not mean colons are bad for human readers, but it does show that for the specific search technology used in this test, the colon acted as a slight obstacle. The researchers noted that this effect was not universal; for many titles, the change made no difference at all, but for a significant portion, the space worked better than the two dots.
The study concludes that while large language models are indeed colonizing scientific titles with a frequency that exceeds human norms, this does not mean they are the sole cause of the recent rise in colon usage. The trend began before these tools were widely adopted, and the increase is likely a mix of changing human habits and machine influence. Furthermore, the presence of a colon is not a reliable way to spot a paper written by a machine, as humans have used them for decades. The real takeaway is that punctuation is not just a stylistic choice; it is a variable that changes how machines read and retrieve information. As artificial intelligence becomes more integrated into the research process, understanding these small shifts in language becomes essential, not just for spotting AI, but for ensuring that the knowledge we create can still be found by the systems designed to share it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.