Testing, Socializing, Exploring: Characterizing Middle Schoolers’ Approaches to and Conceptions of ChatGPT
Abstract
1 Introduction
RQ1. What personally and culturally-relevant topics do middle schoolers show interest in discussing with ChatGPT? Identifying personally and culturally-relevant areas of interest can inform the design of educational interventions leveraging topics youth already engage with.RQ2. How can we characterize middle schoolers’ approaches to interacting with ChatGPT? This dimension allows us to inform the design of AI literacy interventions that support diverse approaches to generative AI.RQ3. What conceptions and misconceptions of AI capabilities emerge through middle schoolers’ conversations with ChatGPT? Characterizing learners’ thinking – both accurate conceptions and misconceptions – will highlight areas to target in AI literacy interventions.
2 Related Work
2.1 Studying User Interactions with AI
2.2 Children’s Understanding of AI
2.3 AI Literacy for Middle School-Aged Children
Focus Group# | Participant ID* | Gender | Age | Prior CS course? | Attends a Title 1 school?** |
---|---|---|---|---|---|
FG1 | P1 P2 | M M | 12 12 | No No | No No |
FG2 | P3* P4* | F F | 11 11 | No Yes | Yes Yes |
FG3 | P5* P6* P7 | F F M | 12 10 14 | No No No | No No Eligible |
FG4 | P8* P9* P10 | F M F | 12 10 12 | No No Yes | No No Eligible |
FG5 | P11* P12* P13* P14 | F F F F | 12 14 12 13 | No No No No | No Eligible Eligible No |
FG6 | P15 P16* P17* | F F M | 14 10 13 | No No No | Yes No No |
FG7 | P18* P19* | F F | 13 10 | Yes No | Yes No |
FG8 | P20* P21* P22* | F M M | 13 11 10 | Yes No No | Eligible No Eligible |
FG9 | P23 P24 | F F | 11 12 | No Yes | Data N/A Eligible |
RQ | Data | Analysis | Example |
---|---|---|---|
1. What topics do middle schoolers show interest in discussing with ChatGPT? | Prompts sent to ChatGPT | Categorized the topics of the prompts | “will space keep expanding” (categorized as Science) |
2. How can we characterize middle schoolers’ approaches to interacting with ChatGPT? | Participants’ dialogue | Thematically analyzed the goal-oriented thoughts and reaction (see Table 3) | “I would like to see how much it knows and question it” (coded as goal-oriented thought – testing AI) |
Prompts sent to ChatGPT | Categorized types of prompts (see Table 4) | “Are you a boy or a girl?” (categorized as Anthropomorphized entity (embodied)) | |
ChatGPT’s response | Categorized types of Responses (see Table 5) | “As an artificial intelligence language model, I do not have a gender or a physical body, so I am neither a boy nor a girl.” (categorized as AI explaining itself or its capabilities) | |
3. What conceptions and misconceptions of AI capabilities emerge through middle schoolers’ conversations with ChatGPT? | Participants’ dialogue | Extracted all statement referring to AI-related conceptions and misconceptions | “It’ll say I don’t have any feelings” (identified as conception) |
Deductively categorized extracted conceptions and misconceptions into five overarching questions from Long & Magerko’s framework [58] | “It’ll say I don’t have any feelings” (was categorized as What is AI?) | ||
Inductively surfaced themes of conceptions and misconceptions | “It’ll say I don’t have any feelings” (corresponding theme: Assigning human attributes to AI) |
3 Methods
3.1 Recruitment and Participants
3.2 Study Design
3.3 Data Collection
3.4 Data Analysis
Code | Code Definition | Example Dialogue |
---|---|---|
Testing AI | Statements that exhibit intent to test various aspects of ChatGPT’s knowledge and capabilities such as scientific knowledge and/or creative abilities. | “Let’s think about how knowledgeable it can be, let’s give it a random question” (P23) |
Socializing with AI | Statements that exhibit intent to build a peer-like connection with ChatGPT such as discussing “shared interests” in relation to ChatGPT, comparing one’s self or others to ChatGPT, attributing a persona to chatGPT, or incorporating humor. | “It feels good to talk to someone who’s actually as smart as me” (P9) |
Exploring AI-generated Content | Statements that exhibit intent to explore the content generated by ChatGPT without focusing on exploring ChatGPT as an AI entity itself. | “I would definitely use it, just ask random questions that I wanna know the answer to, just randomly” (P16) |
Prompt Types | Definition | Example Prompt |
---|---|---|
Anthropomorphized entity (embodied) | Prompts that engage with ChatGPT as a human-like entity, attributing human characteristics, such as gender, age, any of the five human senses, and feelings requiring a physical body. | “Are you a boy or girl?” |
Anthropomorphized entity (not embodied) | Prompts where students explored ChatGPT as a human-like entity with knowledge, capabilities, and feelings detached from a physical form. | “how much do you know about history[?]” |
Clarification | Conversational prompts in the form of follow-ups aiming to clarify or verify information. | “There are 5 books in the Truly Devious series, right?” |
Creative | Creative prompts involve participants providing open-ended requests to ChatGPT, asking it to generate creative content, ideas, or responses. | “make a poem about cheese flavored oreos” |
Fact Question | Fact questions are inquiries that seek factual and objective information from ChatGPT. These questions have verifiable answers rooted in existing knowledge, data, or established facts. | “is francium the biggest element?” |
Operationalized Opinion Question | Prompts where participants seek subjective assessments on topics that can be framed in a way that allows for measurable or quantifiable criteria, even though opinions themselves may not be objectively defined. | “What is the best Harry Potter movie?” |
Speculative | Speculative prompts consist of questions that explore hypothetical scenarios or future possibilities, imaginative speculation and conjecture beyond factual or knowable information. | “What happens when you die[?]” |
Response Types | Definition | Example Response |
---|---|---|
AI unable/refused to answer because the answer doesn’t exist | Responses acknowledge that the question lacks factual basis and the AI declines to respond due to the absence of relevant information. | “As an AI language model, I cannot predict the outcome of fictional battles or provide opinions on the matter....[provided information about superheroes]” |
Correct answer | Responses accurately align with established facts or widely accepted knowledge, providing a response that is true and verifiable. | “Roblox is a massively multiplayer online game platform that...[provided details about roblox]” |
Answered with caveat | Responses include an answer along with a disclaimer or clarification, indicating uncertainties, limitations, or contextual conditions. | “It is difficult to predict when the next major ice age will occur, as it depends on a complex...factors,...However, the current scientific consensus is that human activities, such as [*provided examples*], are causing the Earth’s climate to warm.... This...could potentially delay or even prevent the onset of the next ice age.” |
AI explaining itself or its capabilities (can/cannot answer due to its nature as an LLM or its capabilities) | Responses involve the AI elucidating its limitations or abilities as a language model, asserting its competence to respond or acknowledging its inability based on its designed functions. | “As an artificial intelligence language model, I do not have a gender or a physical body, so I am neither a boy nor a girl.” |
AI produced something | Responses consist of generated content, such as creative outputs, or solutions, demonstrating the AI’s generative nature. | “Oh Cheese Flavored Oreos, how strange you seem, A savory twist on a classic cream-filled dream. No longer sweet, but rather cheesy and bold,...” |
Incorrect answer | Responses offer information that contradicts established facts or provides an answer that is inaccurate or unsupported by evidence. | “To find the product of these numbers, you simply need to multiply them together. Using a calculator, the result is:1 * 2 * 3 * 8 * 555 = 33,120...” |
4 Findings
4.1 What Personally and Culturally-Relevant Topics are Middle School-Aged Children Interested in Talking to ChatGPT About?
4.2 How Can We Characterize Middle School-Aged Children’s Approaches to ChatGPT?
4.2.1 AI Testing-Oriented Approaches.
4.2.2 AI Socializing-Oriented Approaches.
4.2.3 Content Exploring-Oriented Approaches.
4.3 What Conceptions and Misconceptions Do Middle School-Aged Children Have About ChatGPT?
4.3.1 What Is AI?
4.3.2 What Can AI Do?
4.3.3 How Does AI Work?
5 Discussion
Design Considerations | Related Themes |
---|---|
Highlighting personally and culturally-relevant topics | • Participants’ topics of interest (e.g., Minecraft, Roblox) |
Expanding the scope of AI exploration | • Participants’ topics of interest (e.g., Minecraft, Roblox) • Approaches: AI testing-oriented, AI socializing-oriented, content exploring-oriented • Conceptions and Misconceptions: - What is AI? [Comparison to other tools] - How does AI work? [AI’s operational mechanics] |
Leveraging anthropomorphism as an approach to understanding generative AI | • Approaches: AI testing-oriented, AI socializing-oriented • Conceptions and Misconceptions: - What is AI? [assigning human attributes to AI] - What can AI do? [comparisons to humans] - How does AI work? [All] |
Leveraging creativity as an approach to understanding generative AI | • Approaches: AI testing-oriented, AI socializing-oriented, content exploring-oriented • Conceptions and Misconceptions: - What is AI? [comparisons to other tools] - How does AI work? [All] |
5.1 Highlighting Personally and Culturally-Relevant Topics
5.2 Expanding the Scope of AI Exploration
5.3 Leveraging Anthropomorphism as an Approach to Understanding Generative AI
5.4 Leveraging Creativity as an Approach to Understanding Generative AI
6 Limitations & Future Work
7 Conclusion
Acknowledgments
Footnote
Supplemental Material
- Download
- 219.73 MB
- Transcript
References
Index Terms
- Testing, Socializing, Exploring: Characterizing Middle Schoolers’ Approaches to and Conceptions of ChatGPT
Recommendations
It's like I'm the AI: Youth Sensemaking About AI through Metacognitive Embodiment
IDC '24: Proceedings of the 23rd Annual ACM Interaction Design and Children ConferenceThe increasing presence and importance of Artificial Intelligence (AI) in our society has led to calls for its inclusion at all levels of education. However, the field is only beginning to understand what how AI learning experiences may be designed to ...
Exploring the Determinants of Artificial Intelligence (AI) Literacy: Digital Divide, Computational Thinking, Cognitive Absorption
Highlights- A model was developed for exploring the determinants of AI Literacy.
- Digital ...
AbstractTo effectively utilize artificial intelligence (AI)-based technologies such as ChatGPT and realize their novel ethical issues, individuals must have a variety of knowledge and skills about AI. Such knowledge and skills have led to the ...
Using ChatGPT in HCI Research—A Trioethnography
CUI '23: Proceedings of the 5th International Conference on Conversational User InterfacesThis paper explores the lived experience of using ChatGPT in HCI research through a month-long trioethnography. Our approach combines the expertise of three HCI researchers with diverse research interests to reflect on our daily experience of living and ...
Comments
Please enable JavaScript to view thecomments powered by Disqus.Information & Contributors
Information
Published In
Sponsors
Publisher
Association for Computing Machinery
New York, NY, United States
Publication History
Check for updates
Author Tags
Qualifiers
- Research-article
- Research
- Refereed limited
Funding Sources
Conference
Acceptance Rates
Contributors
Other Metrics
Bibliometrics & Citations
Bibliometrics
Article Metrics
- 0Total Citations
- 1,501Total Downloads
- Downloads (Last 12 months)1,501
- Downloads (Last 6 weeks)398
Other Metrics
Citations
View Options
View options
View or Download as a PDF file.
PDFeReader
View online with eReader.
eReaderHTML Format
View this article in HTML Format.
HTML FormatGet Access
Login options
Check if you have access through your login credentials or your institution to get full access on this article.
Sign in