More than half of recent astronomy papers are written with language-model assistance
A hierarchical Bayesian analysis of over 200,000 astronomy papers reveals that more than half of recent astro-ph manuscripts show linguistic traces of language model assistance, a figure vastly exceeding the less than 1% of papers that explicitly disclose such use.
Original paper licensed under CC BY 4.0 (http://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
In the quiet, data-heavy world of astronomy, scientists spend their days translating the light from distant stars and galaxies into stories about the universe. For decades, the written language of these discoveries has been a strict, technical dialect, stripped of flourish and focused entirely on precision. But in late 2022, a new tool entered the scene: advanced computer programs capable of writing fluent, human-like text. These tools, known as language models, quickly became popular for helping researchers draft papers, check grammar, and organize complex ideas. The question that soon arose was not whether these tools were being used, but how widespread their influence had become. If a computer helped write a scientific paper, would the text look different? Would it carry a subtle fingerprint, a specific set of words that a human writer might rarely choose but a machine favors?
A team of researchers at The Ohio State University and the Max Planck Institute for Astronomy set out to find the answer by looking at the actual words in the literature. They analyzed the full text of more than 200,000 astronomy papers submitted between 2015 and mid-2026. Their goal was to count how many of these documents showed signs of computer assistance. They focused on a specific list of words that had become common in machine-generated text, such as "delve," "underscore," "intricate," and "pivotal." These are not technical terms like "redshift" or "supernova"; they are stylistic choices, words that add a certain rhythm to a sentence. Before the rise of these computer tools, astronomers used these words sparingly. The researchers wanted to know if the sudden appearance of these words in recent papers meant that the papers were being written with help.
To get a clear picture, the team did not just count the words; they built a sophisticated statistical model to separate the signal from the noise. They knew that language changes over time, and that some words might become popular for reasons other than computer use. To account for this, they looked at how these words were used in papers written before 2020, a time before the tools were widely available. This gave them a baseline for how often a human astronomer would naturally use these words. They then compared this baseline to the papers written after 2022. Crucially, they also looked at the small number of papers where authors explicitly admitted to using a language model. These admitted papers served as a calibration point, a known reference to help the team understand what a "computer-assisted" paper actually looks like in the data.
The results were striking. By 2025, the researchers estimated that more than half of all new astronomy papers carried a trace of language-model assistance. Specifically, they calculated that 54 percent of the papers showed a vocabulary pattern consistent with computer help, with a margin of error that kept the figure comfortably above 36 percent even under the most conservative assumptions. This number has been rising rapidly; in 2024, the figure was around 20 percent, and by the first half of 2026, it had climbed to nearly 87 percent. This suggests that the use of these tools has moved from a rare experiment to a standard part of the writing process for the majority of astronomers.
However, the story is more complex than a simple count of papers. The researchers found that the "fingerprint" of the computer is becoming harder to see. In the early days of 2023, the papers that used these tools had a very strong excess of the specific marker words—about 3.5 times more than a human-written paper would have. By 2026, that excess had dropped to just 1.5 times the baseline. This happened because authors began to adapt. As the specific words like "delve" became famous as signs of computer use, writers stopped using them so heavily, or edited them out, making the computer's hand less obvious. The tools are still being used, but they are becoming more subtle, blending in more naturally with human writing styles.
Perhaps the most telling finding is the gap between what is happening and what is being said. While the researchers estimate that 54 percent of papers in 2025 were assisted, only 0.81 percent of those papers contained a formal declaration stating that a computer was used. In other words, for every single paper where an author admitted to using the tool, there were about 66 papers that showed the same linguistic traces but said nothing about it. The researchers checked their detection methods carefully and found that they were not missing many declarations; the silence appears to be real. Most of the authors who did speak up said they used the tools for grammar, clarity, or language editing, tasks that many journals already allow. The lack of disclosure seems to stem not from an attempt to hide misconduct, but from a desire to avoid a stigma, as the scientific community has not yet agreed on whether using these tools is acceptable.
The study concludes that the landscape of scientific writing has shifted fundamentally. The tools are no longer a novelty; they are a dominant force in how astronomy papers are produced. The researchers note that this shift is not necessarily a failure of integrity, but a change in practice that the community has not yet caught up with. As the linguistic traces of these tools continue to fade and become harder to detect, the window to measure their impact is closing. The authors suggest that the most urgent step forward is not to ban the tools or to chase a fading signal, but to establish a clear norm of disclosure. If authors simply stated when they used these aids, the field could move forward without the confusion of hidden assistance, turning a practice that is currently ignored into one that is openly understood.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.