Where Knowledge, Kids, and AI Meet
The Story of Wonderwood
I co-founded Wonderwood, a knowledge-building app for curious kids ages 3-10, with Steve Midgley, founder of Learning Tapestry.
Wonderwood is a smarter encyclopedia.
We help kids build knowledge about science, history, people, and technology because when kids know about the world, they are stronger readers. And knowledge snowballs, so the more kids know to start, the more they gain over time.
As a teacher, I worked tirelessly to help kids understand what they read, but I couldn’t give them all the knowledge they needed.
We built Wonderwood to do what schools often can’t.
The Beginning of Wonderwood
In the summer of 2022, Steve shared his belief that AI is a game changer based on his experience with the early days of the internet in the 90s. He suggested that the best way to understand it is to build something with it.
So, we set out to figure out the where the answers to these questions overlapped:
What problems in learning can’t be solved with more human effort?
What does AI do well?
We also wanted to understand AI’s weaknesses so we could design around those potential pitfalls.
I started by reflecting on my experiences as an English language arts teacher, the knowledge and skills children need, and diving into research on learning, reading, language, and cognitive development.
From there, I developed a set of principles that would shape how we designed Wonderwood.
Wonderwood Design Principles
1. Build knowledge essential for future reading and learning.
Strong reading relies on phonics and what kids know. While the “science of reading” movement has emphasized the importance of structured literacy and systematic phonics instruction, knowledge and vocabulary development are equally critical and yet often overlooked. Knowledge comes from processed information and it leads to understanding, or a person’s internal model of the world. Knowledge is also sticky. The more we have to start, the easier it is to gain, and it compounds over time, so it is important to start early. People assume that kids are gaining the knowledge they need, yet schools are cutting back on science and social studies, the subjects that most build topical knowledge. In fact, research shows that increasing time on social studies improves reading comprehension. Wonderwood then is designed to fill this gap.
2. Let kids choose what they want and need to learn.
As a middle school teacher, I saw firsthand how a lack of knowledge hindered comprehension. Many of my students struggled with the texts I wanted them to read, not because they couldn’t decode the words, but because they lacked sufficient knowledge to make meaning from them. To help, I spent countless hours creating experiences for knowledge building.
As a student, my 7th through 10th grade English classes were combined with social studies, so I often tried to replicate that experience. I would create knowledge stations and provide videos and other texts about key topics, but it was all a guess. Even if I had the time to conduct one-on-one interviews with each student before reading each text, I still couldn’t see into their brains to know what they knew or didn’t. The problem wasn’t my effort. No matter how much work I put into trying to figure out what knowledge to help my students build, there was no way that I would be able to help them individually build a wide enough base of knowledge to support them in understanding the texts then or that they may read in the future.
Wonderwood is designed so that children pick the topics, either because they are interested (e.g., hamsters, unicorns, and rockets) or because they are seeking more information (e.g., money, history, and chemical reactions). This agency creates engagement, motivates learning, and supports retention of information for knowledge building.
3. Make it independently usable.
I also thought about how I built knowledge that served me as a reader and learner. For one, I was a voracious reader, but I also sought information independently. When I was a child in the 1980s, we had an encyclopedia from the 1960s. (I’m fairly certain that it didn’t even document that we had gone to the moon!) But, when I wanted to know something, I went there first. It was a source that I, as a child, could access on my own. I really liked the “H” volume because in the human body section there were clear overlays of each of our body’s systems. I’d stack the circulatory and skeletal systems and then teach my stuffed animals all that I had discovered, imagining how it all fit together. I suppose it was my own analog simulation.
Being able to independently explore and seek answers through visuals first (before I could read) and then written text when I could read was formative. These small independent acts of choice and self-directed learning are so important for children, and so with Wonderwood, children can open the app and start learning without an adult, just like I once flipped open the encyclopedia.
4. Ensure that AI-generated content is safe and family friendly.
When I was growing up, the encyclopedia and books at school and from the public library were my main sources of information. Now kids have access to endless content through the internet on smartphones and tablets, but most of it is unsafe. They are almost always two clicks away from graphic or disturbing content. So while we’re in an “Information Age,” kids are still very much left behind. Even seemingly “safe” spaces are often filled with empty, distracting content, such as unboxing videos on YouTube, or addictive video games, or low-quality noise that offers little real knowledge that will support their reading development. We designed Wonderwood to include only content that builds real, transferable knowledge based on family-friendly topics and prompts and that is reviewed by humans for accuracy and quality.
What AI Does Well and Where It Falls Short
With these design parameters, I started thinking about how we might use AI to build knowledge for children. I know that children gain knowledge from three main sources:
Experiences
Conversations with more knowledgeable people
Reading or listening to texts
I ruled out “experiences” early as real-world exploration beats any simulation. Between conversations and texts, the decision was about what AI does well and not, so we began a process to figure that out.
After testing AI for conversations, I saw that it wasn’t ready. Generative tools lack the guardrails to interact safely with children. Knowing this is important and key to why Wonderwood is both safe and accurate. There are many stories of AI products performing poorly, hallucinating or sharing inappropriate information. Given this, we ruled out the direct generation of content between a child and AI that a conversation would require.
Quality Texts
That left reading or listening to texts. I began to explore if AI could help us generate high-quality, accurate texts for kids. We engaged in a series of tryouts and tests with the AI. It would draft a passage, and then I would review the results to determine quality. Eventually, we trained a team to scale the review.
What is our bar for quality texts? Texts need to be accurate and include both topic-specific vocabulary and general academic language that children benefit from encountering in context. They also need to share details that explain the topic and pique a child’s interest in the topic.
I observed that AI performs well when drawing from well-established, factual content. It struggles with nuance, judgment, or obscure topics. That is when it “hallucinates.” But, when used within bounds and reviewed by trained humans, it can create accurate and detail-rich texts, so we targeted our content to topics that are well known, loved by children, and approved by educators and parents. Today we’re seeing near-zero content inaccuracies even before human review.
Quality Images
Our work to determine how to create quality images followed a similar path: test, evaluate, refine. However, getting to quality has taken longer because the AI models are still improving. We even started with one AI model and moved to another because it was better suited to our content and children. One thing that still seems to trip up AI image models is text in the image, but our mantra for working with AI has been: “Skate to where the puck is and score the goal.” The field is evolving fast and models are greatly improving, and we aim to stay just ahead of it and capitalize on the work of experts.
And an unexpected twist? Kids love the weird images. In early testing, my nephew picked the rainbow-colored turtle and the wild-haired dog. He laughed and shared it with his mom. The whimsical, slightly zany images of a dog flying in the air or a person walking on water spark curiosity. They make kids look closer; they notice and think, “Is this real? Can this happen?” For some kids, they also love to find the oddities like a treasure hunt. So we embraced “the zany.” Wonderwood’s aesthetic favors vibrant, cartoon-style images over realistic photos, which was proven in user research with children. We prioritize whimsy, color, and just enough strangeness to delight.
What is our bar for quality images? Images need to contain no blatant inaccuracies, including misspelled or non-English words. The images should relate to the text both in content and context, including being the right art style (e.g., colorful and cartoonish). Lastly, images cannot contain any inappropriate or scary imagery.
Wonderwood: Where Knowledge Lives
Based on all of our learnings, we determined that AI is good at writing texts and creating images for topics that elementary-aged children and younger are interested in learning about. And given my expertise, I know that strong reading is about what kids know. So we created Wonderwood as a children’s knowledge-building app with bite-sized lessons that include images and texts that kids can listen to or read.
Wonderwood has the spirit of an encyclopedia, but it is dynamic and it reads the text aloud. The encyclopedia from my childhood was static. It didn’t respond to my interests, creating a path through the pages. I had to identify and then locate the next topic I wanted to learn about. This is where the strengths of AI begin to show up. AI can generate an almost infinite number of texts, and it can connect texts together in the moment without human intervention. Instead of stopping at the human body overview as I had to in my childhood, Wonderwood will help kids make connections to animal anatomies and medicine and the five senses or the brain. All of that information develops into knowledge and eventually understanding.
And this is just the beginning. We want to learn how the topics children choose and the paths they create in the app to build their knowledge. We’re committed to growing the app as users need more. For example, we want to add expression of knowledge as well as building knowledge. Think interactive learning activities like coloring or doodle pages or simulated play: a child learns about dogs and then draws a picture of the food a dog eats or can take care of a virtual puppy or a child learns about baking and then writes a recipe or makes virtual cookies.
Wonderwood grew from a key problem I had as a teacher: no matter how hard I worked, I couldn’t give each child all the knowledge they needed to thrive. It’s also rooted in the joy I felt as a kid, flipping through the encyclopedia in search of answers that led to even better questions. That early wonder and the thrill of discovering something on my own has stayed with me. Wonderwood is our digital love letter to that kind of curiosity. By blending insights from research with responsible use of AI, we’ve built a dynamic resource where children can pursue knowledge safely and experience wonder-filled learning.
