As large language models (LLMs) reshape how students engage with programming, educators face increasing challenges in maintaining academic integrity (An et al., 2025; McDonald et al., 2025). Drawing on the presenter's doctoral research, this talk explores how visual programming tasks can meaningfully resist unauthorised AI assistance, offering a robust alternative to conventional text-based exercises that are highly susceptible to automation.

Central to this approach is Thonny-py5mode -- a creative coding extension for the beginner-friendly Thonny IDE, developed as part of the broader PhD research project. It enables graphical output via py5, a Processing-inspired Python library for programming interactive graphics, animations, and applications (Schmitz, 2021).

The talk builds on findings by McDanel and Novak (2025), who observed that LLMs often struggle with assignments involving graphical output, especially when correctness hinges on appearance rather than testable outcomes. We present a series of Thonny-py5mode tasks used in undergraduate assessment and evaluate the performance of leading LLMs (GPT-4o, Claude Sonnet 4, Gemini 2.5) in replicating them. This session offers a scalable, discipline-agnostic strategy -- adaptable to other Python graphics libraries -- for promoting conceptual understanding, supporting creativity, and designing resilient assessments in the GenAI era.