Project
Designing for Young Learners
An explorative thesis and workshop on how children aged 6-8 y.o. perceive chatGPT when in a learning situation.
For my thesis during my master’s in digital design, I wanted to explore a technology in a specific context. This, inspired from my internship at LEGO Group, resulted in an exploration of children’s perception of ChatGPT in learning situations. The goal being to bring forward some awareness and recommendations for designers working on future AI-products for children. I approached the project through the lens of research-through-design with a literary review and an explorative study with a group of children.
Overview
Problem
Understanding how 6-8 y-o. children perceive ChatGPT when conducting a task together.
Outcome
Nine recommendations were developed for how designers can develop better AI learning tools for children's learning, based on children's perceptions.
Role
Project manager
UX-researcher
Timeline
Spring 2024
Tools
Condens
Team
Jane Clausen
End result
Recommendations to consider when designing AI-applications for learning situations
Research
Researching perception, children and AI
To better understand the field I had placed myself in I did a literature review of previous research done within CCI (Child-Computer Interaction).
To note is the fact that no research has been done on children and generative AI in this context before.
Litterature search
Limiting myself to research done within CCI, I looked for articles on interaction between children and AI from 2017 and forward in select journals. Focused especially on studies that also tried to understand children's perceptions.
Gather knowledge
Doing a literature review helped with a better understanding of my anchorage and point of departure for my own study. By gathering insights into how previous studies were done I was able to create a test that was both comparable and able to answer the thesis question.
Analysing
The literature found were assessed with a qualitative approach. By doing so, making it possible to develop new theories, narratives or themes that are shared by the different qualitative studies. This was used to help broadening the understanding of children’s perception of AI-agents as well as showing where there is a lack of knowledge.
Thesis overview | stages and phases
While it can look linear the process of researching, testing and analysing was a messy process of untangling, experimenting and finding common grounds.
Insights
Initial findings & uncovered insights
The findings from the literature review helped shape the study I was going to make as well as finalising the thesis question.
Researching perception is not a 'one size fits all'
While researchers goes about this in quite a few different ways. A few findings are shared.
Most children perceive AI-agents' intelligence to be high but they also find children believing the opposite as it lacks emotional intelligence.
They found an increased feeling of friendship between child and AI-agent when children got to learn about the AI-agent.
Geography had an influence on children's perception of AI-agents as their skepticism, trust and general view of the AI-agent was different depending on where in the world the children were from.
A lack of research combining children and generative AI.
A need to understand how generative AI differentiates itself from 'traditional' AI. For this research it means a need to look at chatGPT's ability to hold a conversation, its likeability, intelligence, its levels of humanness, ability to read the needs of the child, and how it being a learning activity impacts the child's perception.
A need for a new language
One study looked at perception through the lens of artefact versus animate, studying whether children perceived the AI-agents to be one or the other. Their results showed children spontaneously attributing AI-agents with both properties. This shows a need for a language to speak of this in-between.
Study design
Remember who the participants are when planning a workshop
I was acutely aware that I was making a workshop for children which meant being conscious of how long they can concentrate, how involved they were going to be, how to make them play along etc.
Participants
15 children ages of 6-8 years old, 9 girls and 6 boys. All were Danish speaking although for a couple it was their second language. They participated in the activity in random groups of 2-3 children.
Activity
The children were asked to create a story together with the speaking version of chatGPT 3.5. As a learning moment and to help support the children in the story making they were also given small notes with illustrations representing different story characteristics on them. This was meant as a visual aid.
Gathering data
Data was gathered through filming and later transcribing the sessions. After the story activity, the session switched over to a semi-structured interview with eight questions pre-prepared inspired by previous studies (from literature review). The interview helped supply context to their actions during the story making.
The setup
Setup's or prototypes for a workshop does not have to be very elaborate to be effective! For this I used my phone with chatGPT and illustrations of story elements the children could use as story guides.
Workshopping with little wild ones
Expectations versus reality, workshop edition
Having a plan of attack when making a workshop is key. Being ready to differ from the plan is just as important.
For the most part the children were amazing, below is only to show the difference in behaviour and unexpected reactions.
Results from study
How children and chatGPT co-create
AI as a creative learning partner
Is it human?
Children primarily saw chatGPT as a tool or helper, but during immersive interactions, they sometimes treated it more like a person. This reflects how easily AI can blur the line between machine and human.
Breakdowns teach adaption
Breakdowns i.e. chatGPT not understanding, mishearing or not listening while processing etc. disrupted every single test. The children were quick to adapt by rephrasing their answers, speaking louder/more clearly and understanding the listen-process-answer cycle that chatGPT follows. Children are very good at understanding and adapting although also quick at deeming the AI as 'not intelligent'.
Us versus it fueled story making
An 'us versus it' dynamic occurred which seemed to fuel creative collaboration between the children. An example could be the children whispering among themselves agreeing on a plot before loudly proclaiming their decision to chatGPT and continuing the story.
Designing for learning
To better support children, AI tools should balance creative freedom with structured scaffolding or guidance. This means tailoring the AI to the specific learning subject, the developmental and linguistic leves and ability to divert conversations back to the subject at hand.
Final thoughts
Reflections on a lively thesis
Researching through design is a balancing act as it is easy to continue the iterative process of learning, reflecting and developing.
Do we ever truly know perceptions?
The balance between AI and parent/teacher involvement.
In the workings of this thesis I kept going back to the thought of 'AI learning tools should not replace parent/teacher involvement in children's learning' and how to balance the benefits of both AI and adults.
Scope and time alignment
Navigating between all the ambitions for this projects and the time I had to do it taught and reinforced the importance of planning and planning for the unforeseen to happen as well.
