Related Experiment Videos
Evaluating the Feasibility of Artificial Intelligence in Generating Visual Abstracts: A Pilot Study
Ramsha Akhund1, Bipul Mainali2, Azeem Izhar1
1Department of Surgery, University of Alabama at Birmingham, Birmingham, Alabama.
The Journal of Surgical Research
|August 5, 2026
Summary
New AI models show promise for creating visual abstracts (VAs) for scientific research. GPT-5.3 demonstrates significant improvements over GPT-4o in generating accurate and clear VAs, offering a useful tool for researchers.
Area of Science:
- Artificial Intelligence
- Scientific Communication
- Medical Informatics
Background:
- Visual abstracts (VAs) enhance scientific research dissemination and comprehension.
- Advancements in AI, particularly language and image generation models, present new opportunities for creating VAs.
Purpose of the Study:
- To evaluate the feasibility of using GPT-4o and GPT-5.3 for generating scientific visual abstracts.
- To compare the performance of these AI models against expert-generated VAs.
Main Methods:
- A pilot study used a dataset of 26 visual and textual abstract pairs.
- AI models were fine-tuned on a training set and tested on a separate set.
- Generated VAs were compared to human-created VAs based on template adherence, accuracy, clarity, and consistency.
Main Results:
- GPT-4o generated VAs with significant deviations, text errors, and poor visual clarity.
- GPT-5.3 demonstrated substantial improvements, producing VAs comparable to human-generated ones in design and consistency.
- GPT-5.3 still had minor errors in specific elements like journal logos and author names.
Conclusions:
- GPT-4o is not yet suitable for generating scientific VAs.
- GPT-5.3 shows potential for creating first drafts of VAs, improving accuracy, clarity, and adherence to templates.
- AI models are becoming increasingly viable tools for scientific knowledge dissemination.