Recently, students in the 乱伦社区鈥檚 noticed a trend on social media: People were using generative artificial intelligence to make short science videos. The trouble was that these people weren鈥檛 scientists, which, given AI鈥檚 proclivity to be convincingly wrong, could accelerate the spread of misinformation. So the lab wondered how to enable scientists and other researchers to better adapt to platforms like TikTok.聽
鈥淭he alternative is that science is being talked about without scientists,鈥� said co-lead author , a UW doctoral student in human centered design and engineering.
Those discussions led the team to build , an AI tool that helps users turn research papers into 45-second videos. A researcher uploads a paper to the tool, which uses Google Gemini to write a short script explaining the paper. The researcher can then iteratively edit the transcript and resulting video clip.
The team April 17 at the Association for Computing Machinery Conference on Human Factors in Computing Systems in Barcelona.
鈥淔or several reasons, most people don鈥檛 read research papers,鈥� said senior author , a UW professor in human centered design and engineering. 鈥淚 still have challenges reading papers in fields I’m not familiar with. So we wanted to find a way to quickly turn papers into a format that laypeople would want to engage with, and we wanted to study how they engaged with it.鈥�
Currently, PaperTok is only accessible to users with a paid Google Gemini subscription. Those users can go to the and upload a research paper. The system then presents four options to use as a hook in the video. For instance, a PaperTok video on PaperTok itself begins, 鈥淓ver get overwhelmed reading a dense academic paper?鈥�
鈥淭o start, we interviewed eight science communicators and content producers about how to make engaging, credible videos,鈥� said co-lead author , a UW doctoral student in human centered design and engineering. 鈥淲e found that hooks are integral to shortform videos. Because you’re competing with other videos online, you have only a few seconds to grab someone鈥檚 attention.鈥澛犅�
After picking a hook, PaperTok generates a script, which users can edit. In the storyboarding phase, the script is broken into scenes 鈥� much like a movie storyboard. Users can keep refining their scripts and video clips. When they鈥檙e happy with the result, they can add a byline, which appears at the end along with the paper鈥檚 authors.聽
The team asked 100 online participants and 18 academic participants to compare video from PaperTok with videos from two other PDF-to-video generators. They found PaperTok easy to use and its videos more engaging than those from the other systems. But some had concerns that it was 鈥渢oo AI-ish鈥� 鈥� because of AI signs like nonsense text 鈥� to want to share publicly, because that may diminish their scholarship鈥檚 credibility.聽
The team plans to keep working on ways to customize the AI-generated video, such as allowing users to draw on specific parts of a scene so that elements change based on their intent.聽
鈥淭he main motivation behind PaperTok was, 鈥楬ow can we enable researchers to create engaging short-form videos?鈥欌�� Cristobal said. 鈥淏ecause with generative AI tools, anyone can generate a video from a PDF in minutes, and that presents all sorts of problems 鈥� misinformation, AI slop. So we wanted to build a tool that keeps humans, ideally experts, involved. If anything, we hope that PaperTok highlights how important people are in science communication.鈥�
Co-authors include, a UW doctoral student in human centered design and engineering; of Boson AI, who contributed to this research as a UW master鈥檚 student;, a UW doctoral candidate in human centered design and engineering;, a UW doctoral student in human centered design and engineering; and, a UW student in computer science. This research was supported by Microsoft AI and the New Future of Work Award, the Google PaliGemma Academic Program GCP Credit Award, and the National Science Foundation CISE Graduate Fellowships.
For more information, contact Hsieh at garyhs@uw.edu, Shin at dhoon@uw.edu and Cristobal at meziah@uw.edu.