Role: Principal investigator
Funder: National Science Foundation, SRI
Team: Cross-functional team of ML researchers and education researchers
Timeline: 2021–present
Project Site
Quick Impact Snapshot:
- 1,000+ videos analyzed with custom rubric
- 13 interviews + 3 focus groups with parents
- 95% model accuracy using parent-centered content definitions
- Findings shared via podcasts, social media, conferences
- Sparked NSF funding and patent application
The Challenge
Young kids spend the majority of their screen time watching videos online on sites like YouTube. Many parents hope kids will learn from the videos they watch online. But it's really hard to tell what's actually educational on YouTube! Parents don’t have easy ways to assess whether a video supports early literacy or math learning. Our team set out to explore a bold idea: Could we use machine learning to help parents identify high-quality learning videos on YouTube?
We weren’t trying to build an app or launch a new platform. Instead, we focused on developing a capability—a backend technology that could eventually power any number of tools, like browser extensions or smart filters. Our goals: define what "educational" means in this context, and build an AI model that could reliably detect it.
Listening to Parents First
We began by asking parents what matters to them. I led 13 interviews and 3 focus groups with diverse parents of preschool and early elementary-aged children. We wanted to understand how they currently chose YouTube content, what they looked for, and where they struggled.
Parents told us they wanted content that was both fun and educational, but they felt overwhelmed trying to determine quality. They were skeptical of YouTube’s recommendation engine and wanted clearer, more trustworthy guidance. These insights grounded our design goals: any system we built would need to reflect real parent priorities and concerns.
We started by understanding the pain points APPROVE might address. For example, one parent explained how meaningful saving time is for her:
"As a single working parent, ten hours a day I'm gone from my house, so I have limited time with my son and I want it to be important and meaningful. So, if there were somewhere I could cut the time of having to do the research on my own, that would be amazing.– Parent of a PreK child
Other parents described frustration with limited parental controls, or a desire for their child to watch more educational content online.
Parents' comments also helped us prioritize APPROVE's features. For example, we initially focused more on content detection than on a parent dashboard because parents said they already monitor what their young child watches online. Nine of the 13 parents in our focus groups said they stay near when their child is watching videos:
"I'm often able to see, or he will show me. It is usually pretty clear waht the video is about, or what the interests of the video is." -Parent of a PreK child
To share these findings with a broader audience, we summarized key takeaways in a blog post: Conversations with Parents About YouTube
Defining "Educational Quality" for Machines—and Humans
Before training a model, we had to define what counted as "educational." I led the development of a coding rubric aligned with national early learning standards—including the Head Start Early Learning Outcomes Framework and Common Core standards for kindergarten and first grade. The rubric also incorporated pedagogical quality dimensions like clarity of instruction, age-appropriate modeling, and language richness. Here are some examples of the educational content codes included in the rubric:
We consulted with experts in early childhood education and children's media to ensure our rubric reflected best practices. Based on their feedback, we expanded the rubric to include not only educational content codes, but also codes for the quality of the instruction. Here's a snapshot of the rubric we created for those codes, with our advisors' input:
Then we trained a team of annotators to apply these codes to over 1,000 real-world YouTube videos. These annotations became the foundation of our model training data.
We described this rubric, the annotation process, and what we learned in a white paper: YouTube’s ABCs and 123s. For broader public audiences, we summarized the findings in two blog posts: one from the Joan Ganz Cooney Center and one from SRI Press.
Partnering with Machine Learning Researchers
With our human-coded dataset in place, I collaborated closely with ML researchers to co-develop the model framework. Our annotations enabled the creation of a multimodal model that analyzed video, audio, and text signals simultaneously.
The model needed to detect subtle distinctions—like the difference between a video introducing rhyming words and one that simply includes them. We worked collaboratively with machine learning researchers, iterating on different solutions for fine-grained content detection, such as providing lists of keywords, reviewing model/annotator discrepancies, and providing feedback on automatic captions. Thanks to our strong collaboration, the final model achieved 95% accuracy in identifying standards-aligned educational content. Here's a diagram of the multimodal content analysis model that powers APPROVE:
We presented the technical contributions of this model in two major conference papers: NeurIPS 2023 (focused on alignment-aware contrastive learning) and CVPR 2023 (focused on handling fine-grained, multi-label classification).
Pitching and Winning Funding
This project started with an idea I pitched at SRI’s internal Shark Tank competition—and won. That $100K award helped us prove early feasibility, attract internal co-investment, and eventually secure a $300K grant from the National Science Foundation. We're now applying for additional funding from NICHD to expand this work further.
"Given the wide range of free content, this tool could be an economical solution to providing preschool-aged children from low-income communities quality educational content. Additionally, the dataset would be quite large, allowing for important insights into preferences and educational screening impacts." -NSF Review Panel
More on the NSF-funded portion of the work: NSF Award #2139219
Sharing What We Learned
We’ve shared our work with academic and public audiences alike:
- Conference presentations at SRCD and Digital Media & Developing Minds
- Podcast appearance discussing project findings: SRI Homeroom Podcast: What Are Children Watching Online?
- Social media video viewed over 4,500 times and used to increase public awareness of the project
What's Next for APPROVE
We'd love to expand APPROVE to detect more educational topics. In focus groups, parents identified a wide range of educational topics on which they'd like their kids to watch YouTube videos, including music and art and how-to (e.g., cooking, drawing). We also plan to expand APPROVE to analyze content on other video viewing platforms, based on parent feedback.
Reflections
This project brought together some of my favorite things: elevating parent voices, translating learning science into usable tools, and bridging research and technical innovation. I'm proud that our work has laid the groundwork for future tools that could help families navigate the vast world of kids’ media—and make learning a little easier to find.