Demystifying the oracle: A "20 Questions" game to promote AI ethics and literacy

LLMs are not fact-retrieval systems: The paper demonstrates that students commonly misunderstand Large Language Models as "oracles" that access a fixed internal database of facts. In reality, LLMs operate as probabilistic engines that generate responses based on statistical patterns, not stored information—a fundamental distinction for responsible AI use.

Interactive experimentation reveals AI limitations: Through a modified "20 Questions" game, students discovered that identical conversation histories can produce different answers from the same model. This hands-on observation of the model's stochastic (probabilistic) nature helps users develop "informed skepticism" rather than blind trust in AI outputs.

Verification of AI outputs is a technical necessity, not just a rule: By framing output verification through physics concepts and observable model behavior, the paper shows that checking AI-generated information becomes understandable as a requirement stemming from how these systems actually work, rather than an arbitrary compliance requirement.

Anthropomorphization undermines responsible AI use: The paper identifies a critical ethical challenge in education: students' tendency to treat LLMs as intelligent authorities. This misunderstanding can lead to over-reliance on AI outputs without appropriate critical evaluation, highlighting the need for better AI literacy education.

Familiar concepts can scaffold AI understanding: Mapping LLM behaviors to established physics principles provides an effective educational framework for non-technical audiences to grasp how generative AI actually functions, supporting more informed and responsible interaction with these systems.