Thursday 06 March 2025
Scientists have long recognized the importance of captions in scientific publications, as they help readers understand complex figures and data visualizations. However, many captions are poorly crafted, often due to a lack of attention from paper authors. A recent study has shed light on how paper writers use AI-generated captions in their writing process.
The researchers conducted a user study involving 18 participants who rewrote captions for two figures from their own recently published work using AI-generated captions as a resource. By analyzing video recordings of the writing process, the team observed that participants often began by copying and refining AI-generated captions. This suggests that paper writers are willing to use AI-generated captions as a starting point, but also want to tailor them to their specific needs.
The study found that paper writers favored longer, detail-rich captions that integrated textual and visual elements. They also noted that current AI models were less effective for complex figures, highlighting the need for more sophisticated AI systems that can better support the challenges of academic writing.
One interesting aspect of the study was the way participants interacted with the AI-generated captions. The team identified three main integration activities: copy-retain, adapt, and delete. Copy-retain involved using the AI suggestion in its original form; adapt involved incorporating the core concept while modifying linguistic expression; and delete involved discarding the AI suggestion altogether.
The study also provided insight into how paper writers used different configurations of GPT-4o, a large language model, to generate captions. The results showed that the model performed well when generating captions with a 30-word limit, but struggled with longer captions. This highlights the need for more research on developing AI models that can effectively generate high-quality captions.
The findings of this study have significant implications for the scientific community. By better understanding how paper writers use AI-generated captions, researchers can develop more effective tools and systems to support academic writing. For example, AI models could be designed to generate longer, detail-rich captions that are tailored to specific figures or data visualizations.
Moreover, the study suggests that paper writers may benefit from using AI-generated captions as a starting point, rather than trying to write them from scratch. This could save time and effort, allowing researchers to focus on more complex aspects of their work.
Ultimately, this study highlights the importance of understanding how humans interact with AI systems in writing tasks.
Cite this article: “Human-AI Interaction in Scientific Caption Writing: A Study on Paper Writers Use of AI-Generated Captions”, The Science Archive, 2025.
Ai-Generated Captions, Scientific Publications, Figure Captions, Data Visualization, Academic Writing, Paper Writers, Gpt-4O, Language Model, Caption Generation, Research Tools
