Collaborative Research: CER: Educating the Next Generation for Copyright and Privacy Risks in LLMs

Artificial intelligence (AI) tools that generate text and images, known as large language models (LLMs), are now widely used in classrooms and workplaces. These LLMs can unexpectedly reproduce copyrighted materials from books and other protected works, or reveal private personal information contained in the data used to train them. The same LLM system may refuse a request in one instance yet reveal protected content in another, making it difficult for students, teachers, and other users to determine when these tools can be trusted. Recent lawsuits and major settlements involving the use of copyrighted materials demonstrate the significance of these risks. Despite these concerns, most computing and cybersecurity courses treat AI as a finished product to use rather than a system to be scrutinized and tested. As a result, students rarely learn how to examine AI behavior, identify potential vulnerabilities, and assess risks. This project addresses that gap by providing teaching materials, practical tools, and a website that enable learners to investigate how AI systems handle copyrighted and private information, collect evidence, and communicate their findings. The project serves the national interest by helping prepare a skilled workforce for fast-growing jobs in AI safety, oversight, and security. Because the website and tools run in a standard web browser and require neither special equipment nor programming experience, they can be adopted by K-12 schools and under-resourced institutions, expanding access to AI and cybersecurity education. This project researches an education-centered pipeline that teaches high school and undergraduate learners to detect, measure, and communicate copyright and privacy risks in LLMs, and develops a use-inspired theory for AI security education tools. The work is organized into three interconnected activities. First, the project develops a flexible curriculum centered on AI safety and red teaming -- the practice of deliberately testing a system to identify and expose its weaknesses. Offered as a full course, a shorter module, or a high school version, the curriculum trains learners to probe AI models, assess similarity to protected sources, estimate uncertainty, interpret findings within legal and ethical frameworks, and communicate evidence clearly and effectively. Second, the project builds Copyright Detective, an open-source, web-based tool offering separate tracks for computing students and the public that supports controlled prompting of AI models, visual comparison against reference text, multiple similarity measures, and uncertainty estimates across runs. Building on a working prototype that detects word-for-word reproduction, the project adds components for memorized knowledge, persuasion-style jailbreaks, and drift across model versions. Third, mixed-methods studies examine how learners using Copyright Detective reason about uncertain, context-dependent system behavior; the findings from these studies will inform iterative refinement of both the curriculum and Copyright Detective. These activities produce new teaching knowledge, portable low-barrier resources, and an openly released platform that advances workforce readiness for AI security, privacy, and trustworthiness. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.