About this paper symposium

---

Panel information
Panel 24. Technology, Media & Child Development

Paper #1
Learning with AI or Humans: Children's Perception of Mind in AI Learning Companions
Author information	Role
Dr. Ying Xu, Ph.D., Harvard Graduate School of Education, United States of America	Presenting author
Trisha Thomas, Harvard Graduate School of Education, United States of America	Non-presenting author
Chi-Lin Yu, Department of Psychology, Oklahoma State University, United States of America	Non-presenting author
Abstract
Children acquire much knowledge through asking questions and talking with others. With AI technologies like Alexa and ChatGPT also having “conversations” with children, these systems may provide additional learning opportunities (Wang et al., 2024). However, given that many AI learning companions share human-like characteristics, children’s ability to distinguish between human and AI interlocutors poses a challenge with potential implications for their learning and social development. This study investigates whether children aged 4-8 can differentiate between AI and human conversational partners using a Turing Test-inspired paradigm. We hypothesized that children rely on linguistic cues but may misclassify based on appearance, positioning AI and present humans at opposite ends of a gradient, with hidden humans in between. A total of 119 children (66 female, M = 6.52 years, SD = 1.38) participated in joint storytelling activities across three conditions. In the AI condition, children interacted with a generative AI chatbot (GPT-4) through a smart speaker. In the hidden human condition, children interacted with a human partner through the same smart speaker. Children also interacted with a physically present human experimenter, serving as a baseline. Children’s mind perception was measured using scales for agency, experience, and perceived homophily (Severson & Lemm, 2016). Communicative behaviors were assessed through the quantity and quality of responses. Results revealed a linear increase in perceived human-like attributes from AI to hidden human to present human, supporting the gradient hypothesis. Children attributed the highest levels of agency and experience to present humans, intermediate levels to hidden humans, and the lowest levels to AI. In the hidden human condition, their perceptions of agency and experience were similar to those of the AI, with Bayesian analyses suggesting subtle differences. Although children exhibited different communication behaviors with present humans, these differences diminished when controlling for speech characteristics, indicating that linguistic cues, rather than mind perception alone, may drive their communicative behaviors. This supports the idea that human-like linguistic cues can trigger socially ingrained responses, even if children do not consciously perceive their partner as human. Overall, our findings highlight the importance of physical presence and linguistic cues in shaping children’s perceptions and interactions with AI and humans. The linear increase in perceived attributes underscores the gradient nature of children's perceptions. Additionally, findings suggest that older children are more discerning in attributing human-like qualities, particularly experience, to present humans. This study shows that children’s ability to distinguish between AI and humans becomes challenging when external cues are limited. These findings encourage further exploration of how children’s social cognition and behaviors interact when engaging with AI and how this may impact child development. They also underscore the importance of considering how AI technologies might shape children's learning experiences. As AI systems become more integrated into educational settings, it is vital to understand whether children perceive these agents as capable of human-like thought and emotion. Understanding the interplay of children’s perceptions and behaviors during AI interaction can inform the design of educational technologies that enhance learning and promote healthy social and cognitive development.

Paper #2
Adolescent cheating with generative AI after a full-academic year: A status check
Author information	Role
Victor R. Lee, Stanford University, United States of America	Presenting author
Ruishi Chen, Stanford University, United States of America	Non-presenting author
Annie Camey Kuo, Stanford University, United States of America	Non-presenting author
Denise Pope, Stanford University, United States of America	Non-presenting author
Sarah Miles, Challenge Success, United States of America	Non-presenting author
Abstract
ChatGPT and other generative AI chatbot technologies have complicated our notions of screen-based interactions that children will have. For education, there was a fear of cheating because of the human-like text it could produce. In US schools, the response has varied from initial complete bans to adopting the tool as a resource in the classroom (Elsen-Rooney, 2023; Yan et al., 2023). Research on the 2022-23 school year had come out that surprisingly showed no significant increase in cheating that first year relative to prior years (Lee et al., 2024). The reasons for this were complicated - on the one hand, cheating was already persistently high and on the other, there was actually quite a bit of nuance regarding what students thought was acceptable to do with AI. A full year later, we ask - whether the Lee et al., (2024) result was an anomaly. Following from that study, we check what types of school-related tasks have high school students used AI chatbot technology for and have they changed between 2022-2023 and 2023-2024? To what extent did high school students’ ideas about what should be permitted with AI change? In the 2022-2023 study, one charter, one private, and one public were studied. For this session, we expand to two of each school type, with five new schools. In total, have we report on 2242 responses from high school students across six schools. The survey was administered in the spring of 2024. For the preliminary analysis, descriptive statistics were examined for each school. Additional analysis includes non-parametric testing (e.g., chi-squared) on subgroups similar to Lee et al. (2024). When examining what types of school related tasks high school students have used chatbot technology for in the last month (2023-2024), the highest response for five schools (77-84%) was “to explain concepts, problems, or topic.” The second highest response for five schools (57-81%) was to generate ideas for a paper or project. These results were similar to the 2022-2023 (Lee et al., 2024), where the two highest responses of AI usage was also “to generate ideas for a paper or project” at one school (50%) and “to explain concepts, problems, or topic” at another school (43%). As will be discussed in the session, the six schools in the 2023-2024 study continue to show similarity to the 2022-2023 findings. Students favored chatbot usage for tasks such as explaining concepts, but not for writing their papers for them. The immediate fear with the release of generative AI chatbot technology was that adolescent cheating would increase, specifically with writing. Lee et al. (2024) suggested that was not the case, and this follow-up study with an expanded sample and longer time with AI availability suggest those findings are holding still. The ways in which screens full of generative AI impact impact learning experiences will change. However, the evidence suggests we need a more nuanced look at how children are perceiving and using this type of screen-based interaction.

Paper #3
Do Educators Prefer GenAI Educational Apps?
Author information	Role
Dr. Adam Kenneth Dubé, Ph.D., McGill University, Canada	Presenting author
Emma Liptrot, McGill University, Canada	Non-presenting author
Abstract
There is a dramatic rise in educational apps incorporating AI in recent years (Li et al., 2022), which is further accelerated by advances in generative AI (GenAI) and the introduction of OpenAI’s app store in 2024. Educators already struggle to select good apps from app stores, which are oversaturated with poor-quality options (e.g., Dubé et al., 2020; Papadakis et al., 2018). To help sort through an increasing supply of GenAI educational apps, it is essential to understand how GenAI impacts educators’ choices. Educators look for five educational benchmarks to choose apps for their elementary classrooms; curriculum, feedback, scaffolding, learning theory, and development team (Liptrot et al., 2024; Montazami et al., 2022). Educational apps can use AI to deliver curriculum (e.g., Leong et al., 2024), provide feedback (e.g., Colliot et al., 2024), provide adaptive scaffolding (e.g., Callaghan & Reich, 2021; Chen et al., 2023; Lim et al., 2023) or guide their instructional approach (i.e., learning theory). However, it is unclear whether educators prefer apps that use GenAI over non-AI apps to meet these benchmarks. This study tests whether the presence of GenAI impacts educators’ choice of educational apps (i.e., AI present: Yes, no) and if educators prefer specific AI uses (i.e., curriculum, etc.). No hypotheses regarding direction of differences were made (i.e., AI > no AI). 101 educators of K-6 students in North America (28.7%) and the United Kingdom (71.3%) were recruited through the online platform Prolific (74% Woman, 77% white, 41% urban, 83% public). Participants viewed 8 images of fake educational math app pages (figure 1). All app pages mentioned the five benchmarks in their descriptions, but half mentioned using GenAI to either 1) generate curriculum content, 2) provide scaffolding, 3) provide feedback; or 4) guide learning theory, while the other half did not mention AI. After viewing each app, educators evaluated it by stating 1) how likely they would be to download it, 2) what they would rate them, and 3) how much they would pay for them. A repeated-measures MANOVA revealed no main effect of GenAI on educators’ evaluations, Λ=0.95, F(3,98)=1.30, p=0.279, η²p=0.04. However, a MANOVA on the four AI apps found a main effect of use, Λ=0.53, F(9,92)=9.12, p<.001, η²p=0.47, with follow-up ANOVAs revealing significantly lower evaluations for AI Feedback (ps <.001). Furthermore, custom contrasts revealed significant differences between individual AI-uses and the average non-AI app (e.g., AI Feedback < Non-AI: AI Scaffolding > Non-AI; ps<.01). See table 1 for means. Preliminary results provide no evidence that educators have an overall preference for whether an app uses GenAI. Researchers should take advantage of the opportunity to develop AI-specific app selection guidelines that will inform educators’ future preferences. Further, educators seem to prefer apps with AI Feedback less than other AI apps and may like AI scaffolding more than apps without AI. Results of our continuing data collection and analyses will further detail whether educators prefer GenAI for specific benchmarks and whether previous use of AI in their teaching impacts their preferences.

⇦ Back to session