"It's smart and it's stupid:" Youth's conflicting perspectives on LLMs' language comprehension and ethics

While experts disagree on whether large pretrained language models (LLMs) understand human language, youth use, play, and experiment with LLM-powered applications every day. Yet little attention has been given to young people’s perspectives on LLMs’ capacity for understanding and the ethical implications of synthetic texts produced by these systems. We conducted a participatory design session using big paper methods with 18 14-15-year-olds in which they (1) interacted with ChatGPT (GPT 3.5) and (2) prompted it with commonsense questions. We examined youths’ big papers, prompts, and conversations during the activity. Our analysis shows that youth had conflicting views on GPTs’ language comprehension and task completion, as well as synthetic text’s ethical implications with regards to misinformation, privacy, and safety. We discuss how these findings could inform the design of learning activities.