Kids watch a lot of online videos. SRI is exploring how machine learning could steer that screen time toward positive learning outcomes.
In a major 2020 survey, 80% of parents in the United States said that their child watches YouTube. A more recent 2024 survey found that 30% of kids watch more than two hours of YouTube and YouTube Shorts per day.
How much of that content is educational in any meaningful way, and what might kids be learning from it? That is a crucial but largely unanswered question. Researchers at SRI have turned to machine learning to help solve the puzzle.
SRI’s APPROVE (Assisting Parents to Review Online Videos for Education) project is developing an AI-powered app to help parents and teachers cultivate healthy media use while also advancing our knowledge of how online videos impact kids.
How do you classify an endless sea of content?
One of the biggest challenges of research on online media for kids, explains SRI senior education researcher Claire Christensen, is the sheer volume of content. According to YouTube’s own statistics, users upload an average of 20 million videos each day. At this scale, manual classification, in which humans watch videos and code their content, is impossible. So the APPROVE research team made an AI tool to do it instead.
“Our goal is to help increase young children’s exposure to high-quality educational media and give parents and educators better tools to support learning in the digital age.” — Claire Christensen
APPROVE is a machine learning model that automatically detects early literacy and math content and teaching quality in videos. APPROVE detects preschool and kindergarten math and literacy content aligned to the Common Core State Standards and the Head Start Early Learning Outcomes Framework. It also detects how videos teach these topics — the quality of the instruction. For example, does the video ask children questions and wait for their response? Does it use characters and stories to teach?
Developed by an interdisciplinary team of children’s media researchers, early learning researchers, and machine learning experts and trained on a dataset of nearly 1,200 human-annotated videos, this model leverages multimodal content detection to process audio and visual signals in videos.
A new peer-reviewed paper, based on work funded by the National Science Foundation, provides some encouraging early data on the performance of the APPROVE model. Among the findings: The model can detect early math content with 92% accuracy, and can detect indicators of the quality of teaching with up to 95% accuracy.
The next phase of research: fast-paced content and other persuasive design
The team is also building APPROVE to detect indicators of persuasive design — production elements intended to capture children’s interest, such as fast pace, bright colors, and sound effects. One particular area of interest for the APPROVE team is fast-paced content. Multiple previous studies have shown an association between exposure to fast-paced videos and adverse outcomes among children aged 8 and under, including diminished executive function and attention. Fortunately, in a new white paper, the APPROVE team describes a temporal machine learning model that can detect fast-paced content with 85% accuracy and that has potential to save time and labor over typical human annotation methods.
“Our goal is to expand this work to detect additional persuasive design features, such as wish fulfillment themes, bright colors, loud sound effects, and intense emotion, and to link these features to developmental outcomes,” explains Christensen. “Our longer-term goal is to integrate these models into a longitudinal study tracking children’s media use and its impact on attention, executive function, and problematic media use.”
From the lab to the home and classroom
To help children, parents, and teachers navigate the ocean of online video, it’s critical that we better understand and identify the specific features that shape developmental outcomes. The APPROVE project has already proven that it is possible to detect those features at scale, and the model’s accuracy will continue to improve as the researchers train the model on larger, more diverse datasets of videos watched by real kids. The team also aims to adapt the APPROVE model to detect standards-aligned content in other subjects, like science, and at other age or grade levels.
In the long run, the vision for APPROVE includes a user-friendly application that helps parents and teachers curate educational media for kids. Imagine a parent filtering recommend videos based on educational content (and getting AI-generated conversation prompts to reinforce learning) or a teacher using APPROVE to save time searching for the perfect instructional media.
“Ultimately,” Christensen concludes, “our goal is to help increase young children’s exposure to high-quality educational media and give parents and educators better tools to support learning in the digital age.”
Learn more about SRI’s digital learning and technology program.
