This poster examines a workshop held at a local science museum during a spring and summer camp. The 16 consenting participants in the program consisted of 9th-grade students (ages 14-15) from the Philadelphia area. The workshop included activities to promote discussion about electronic textiles, machine learning, harmful bias, and auditing. The participants learned to use tools like Makecode and micro:bits, Google’s Teachable Machine, and more. In one of the workshop days, we explored how youth perceive and understand large language models (LLMs) like Chat GPT.

In order to collect student ideas about Chat GPT, we coordinated small group and classroom discussions. We prompted youths with questions such as: “What are things you already know about Chat GPT?”; “How do you think Chat GPT works?”; “What are some situations in which Chat GPT doesn’t work?”; and “How should people use Chat GPT?”. We tasked the participants to use Chat GPT to answer queries they may have; for example, having Chat GPT complete writing assignments or roleplay a character. We also asked youths to prompt Chat GPT with simple logic questions such as “What weighs more: 2 pounds of feathers or 1 pound of rice?” 

After analyzing the data, we came to several conclusions. We first saw that the students had varying levels of knowledge about how Chat GPT is trained and generates responses. Generally, the participants agreed that Chat GPT used prior data derived from the Internet. However, none mentioned the specific steps to training the Chat GPT algorithm and could not explain how Chat GPT responded to prompts. Another observation was that the youths saw Chat GPT as a helpful tool for writing and brainstorming but noted that it had limitations. Chat GPT had no access to personal information and could not answer certain moral questions. Lastly, the participants disagreed about whether Chat GPT had morals and emotions or if it could be used to cheat. The students debated whether or not using Chat GPT to write an essay assignment was considered cheating.

We encourage more exploration of youth interaction with AI/ML models like Chat GPT. More specifically, it would be fruitful to explore different AI/ML models and how students interact with them. For example, the AI image generator “DALL-E” is an excellent visual tool that enables students to observe examples of harmful bias. Future work should also expand this investigation to a K-12 classroom environment as opposed to a summer camp dedicated to students passionate about STEM. Overall, it’s necessary to explore the use of Chat GPT and machine learning in classroom activities, such as integrating it within a computer science course that covers AI. How could this be implemented in a machine learning course in high school? Or perhaps in an introductory computing course in middle school?