The Challenge

Children and AI

Understanding how machines learn is critical for children to develop useful mental models for exploring artificial intelligence (AI) and smart devices that they now frequently interact with. However, these smart devices are designed to be like black boxes for ease of use and rapid adoption. As a result, children often have limited knowledge of how they work. As AI systems become more integrated, it is imperative for AI curriculum be incorporated into standard 21st century computing education. This project explores new boundaries for K-12 AI education and aims to teach children about AI and machine learning (ML) concepts.

Our goal for the project was to introduce a new, non-programming approach to teaching AI and ML. We leverage children’s familiarity with smart voice assistants to create an experience using the modality of voice. Our ambitions were to demonstrate a conversational experience that was both accessible and effective in teaching children about AI and ML.

Our high level goals were to:

1. Make it accessible and intuitive to use for children.
2. Demonstrate the effectiveness of a voice interface in education of a new field.
3. Create a platform for innovation and engagement.

My Role

I led the design of the learning experience and curriculum between March and August 2019 and collaborated directly with two engineers. We worked closely on the design and content in a rapidly iterative manner.

We wrote and published an academic paper where I am first author, detailing the design, technology, and user study of this project. The final design of this project is currently in development.

‍

The Process

Discovery

This project required heavy literature review on two fronts: the first, on the current landscape of AI in education, and the second, on how to create an intuitive conversational interface for effective learning. 

AI in K-12 education has been evolving and expanding. It dates back to work by Seymour Papert and Cynthia Solomon using LOGO programming and the Turtle robot. Since then, there have been many platforms using block-based programming environments and the context of robotics to teach AI and ML. Conversational agents have also been used in education, from simple text-based input outputs to embodied agents. They utilize natural language dialogue to provide students with a social environment that can motivate learning. 

Before even thinking about design, it was important to define success metrics and understand whether our approach was a correct one. In particular, we wanted to test whether learning happens and how engaged the students were. We used the framework below and a mixed-method approach to investigate our system.

‍

Introducing Zhorai

Zhorai is a teachable conversational agent (and an alien visiting Earth!) designed to teach AI and ML. Students go through four online modules where they (1) analyze Zhorai’s knowledge representation (2) teach Zhorai about animals (3) witness Zhorai learn and guess which ecosystem an animal lives and (4) discuss the ethics of Zhorai. 

“I want to teach Zhorai more. I want to teach Zhorai things like my favorite color is blue.” (student testimonial)

Three primary questions informed my design strategy:

1. How do you design for conversation in an accessible way?
2. What contexts need to be considered?
3. What is effective learning in AI and ML?

It was important to understand the different factors that may influence the student experience early on. Because the interaction is so novel and complex, I mapped all the possible concepts and translated this into the spectrums and situations framework (Adopted from Simon Pan's Uber case study -- I thought it was a great way to organize contexts).

‍

In designing for education, I tried to highlight the range in contexts learners tend to be in, both permanent and temporary, as well as understand the challenges that may occur during learning. 

Supporting Recognition without Recall

One thing we paid special attention to is cognitive load and effort. A voice interface can only be effective if the user does not have a high cognitive load. In an environment where attention is limited, our design supports recognition, and not recall. Using visualizations to show what the agent has learned is opening the black box so that children can trial and error and understand how the agent is learning. We tested the designs of these visualizations to ensure children were able to draw correct conclusions about the agent.

Grade-Appropriate Curriculum

While most AI curriculum is geared towards middle to high school students, our curriculum targeted a younger grade band: 3rd-5th graders. Designing for a lower grade band defined several approaches to our design. First, we created a narrative to capture interest. Students were teaching an alien visiting Earth about Earth’s ecosystems. During user sessions, students were very enthusiastic once they were introduced with this narrative. We also paid special attention to adding an AI Ethics module. Not only do we want students to understand how machines learn, but also the real-world implications of AI in society.

Designing for Trial and Error

The design supports students to go back and retrain the agent to see how manipulating the input changes the resulting output. Students may delete sentences and have Zhorai learn new sentences.

How We Got There

Prototyping the Interaction

Conversation between the student and agent is core to the learning experience. Our early prototype required a facilitator to guide the activity, as well as lead the discussion on ethics.

We envisioned sample dialogue between the student and the agent every part of the activity:

In the beginning, students go through an introduction flow to get to know Zhorai.

The activity flow below illustrates a sample of the intents, slots, and overall conversation:

‍

Feature Prioritization

With the overarching dialogue set, we sat down together to prioritize what features to build.

Iterating on the Design

Ultimately, the system and curriculum were designed to teach children about machine learning. We accomplished this with a learning experience that first makes students think, “Wait, why did it do that?”, then invites them to interact and investigate the underlying model, and finally allows them to make connections and experience success. The final deliverable was a web and tablet-friendly experience, catered specifically towards the demographic of children 3rd grade and up.

Unexpected Challenges

Designing a voice interface comes with its challenges. The inaccuracy of the speech recognition led to micro-frustrations from the student. Allowing the student to delete accident sentences alleviated some of this pain point. 

In addition, children are wildly imaginative and taught Zhorai things about ecosystems and animals that were very interesting, though sometimes irrelevant to characteristics that define them. As the agent is being trained in real time using a relatively small dataset, it could cause the agent to guess incorrectly multiple times. While students enjoyed the trial and error process, it was also important for them to experience success in the end.

Outcomes

From our testing and subsequent user sessions, we found students to be extremely engaged with Zhorai, with groups giving an average of 10 training examples per animal, and individuals giving an average of 12 utterances throughout the modules. This is a significant improvement as children tend to tire quickly using text-based machine learning systems. 

In terms of understanding, children who interacted a lot with Zhorai had a better understanding of how machines learn (based on a worksheet assessment). Overall, children enjoyed interacting with Zhorai and were excited to teach Zhorai more things. Some students said, “I want to teach Zhorai more. I want to teach Zhorai things like my favorite color is blue.” Children were also very immersed in the ethics discussion, touching on important themes such as over-reliance on AI and how to reduce harm, saying things like, “In the future, humans will start depending and relying on AI a lot more, then they could easily make a mistake and not realize that the person they’re trying to help is not getting helped.”

‍