PaperTok: Exploring the Use of Generative AI for Creating Short-form Videos for Research Communication

Paper Title

PaperTok: Exploring the Use of Generative AI for Creating Short-form Videos for Research Communication

Publication Info

• Topic area: Generative AI applications in science communication
• Keywords: generative AI, science communication, short-form videos, human-AI collaboration, large language models, text-to-video, research dissemination, AI credibility, public engagement, participatory web

Background and Problem

• Problem / challenge: Researchers face significant barriers in creating engaging short-form video content due to time constraints, lack of technical skills, and challenges in translating academic findings into accessible formats.
• Significance: Short-form videos on platforms like TikTok and Instagram Reels are increasingly dominant in public information consumption, offering a critical opportunity for science communication to reach broader audiences.
• Motivation and related work: Prior efforts in science communication have leveraged blogs, podcasts, and social media, but short-form video remains underexplored. Existing PDF-to-video platforms lack collaborative input and often fail to balance engagement with credibility. Generative AI offers potential solutions but raises concerns about accuracy, trust, and misuse.

Solution

• Proposed approach: PaperTok, a human-AI collaborative system that transforms academic papers into short-form videos through iterative workflows combining AI-generated scripts, visuals, and voiceovers with human refinement.
• Novelty:
  1. A structured pipeline for generating engaging hooks and scripts tailored to short-form video formats.
  2. Integration of text-to-video models with human-in-the-loop editing for iterative refinement.
  3. Credibility-enhancing features like attribution screens and researcher sign-offs.
  4. Empirical evaluation of AI-generated videos compared to existing PDF-to-video platforms.
• Procedure and key techniques:
  • Researchers upload a PDF of their paper.
  • AI generates multiple hook and script options, which users can refine.
  • Scripts are segmented into scenes, paired with AI-generated visuals, and iteratively edited.
  • A credit screen with authorship attribution is appended to enhance credibility.
  • Final videos are merged and made available for download.

Results

• Concrete findings:
  • PaperTok videos were rated highest for engagement (M = 3.91), informativeness (M = 4.09), and believability (M = 3.92) compared to SciSpace and PDFtoBrainrot.
  • Researchers found PaperTok intuitive and effective for lowering barriers to science communication, though visual inconsistencies and AI artifacts were noted as limitations.
• Advantage over baselines: PaperTok videos were significantly more engaging and entertaining than SciSpace and PDFtoBrainrot while maintaining comparable levels of informational value.
• Experiments / evaluation:
  • Mixed-methods study with 18 researchers and 100 audience participants.
  • Participants evaluated videos across 11 dimensions, including engagement, accuracy, and production quality.
  • Researchers used PaperTok to create videos of their own papers and provided feedback on usability and output quality.
• Limitations and future work:
  • Current text-to-video models produce inconsistent visuals, limiting professional use.
  • Researchers desire more precise control over AI-generated assets.
  • Future iterations should explore domain-specific applications, improve visual generation, and study videos’ impact in real-world dissemination.

Summary

PaperTok demonstrates the potential of generative AI to support researchers in creating engaging short-form videos for science communication. Through a human-AI collaborative workflow, the system balances efficiency with researcher oversight, generating videos that outperform existing platforms in engagement and informational value. However, challenges in visual consistency and researcher control highlight areas for improvement. PaperTok contributes to the broader conversation on responsible AI use in science communication, emphasizing human-in-the-loop design and credibility signals to foster trust and accuracy.