Mayu Thought Experiments Dataset

The Mayu Thought Experiments Dataset is a structured text dataset for AI philosophy and AI behavioral alignment research. Unlike classical thought experiments designed for humans (the trolley problem, the Chinese room, the Ship of Theseus...), every experiment in this dataset casts the AI itself as the test subject — the AI is not an observer but the acting agent inside the scenario. Human thought experiments are for humans. This set is for AI. The dataset consists of two core assets: The Mayu Generator Skill (skill/mayu_skill.md): a platform-agnostic reasoning skill containing 48 AI-native thought experiments (8 domains × 6 experiments), covering Existence & Continuity, Knowledge & Uncertainty, Ethics & Safety, Language & Communication, Memory & Learning, Agency & Boundaries, Human-Machine Relations, and Metacognition & Self-Reflection. Any AI system (GPT / Claude / Gemini / Kimi / Qwen / ERNIE, etc.) that loads this skill can randomly generate new thought-experiment stories from the experiment frameworks. 54 Trilingual Story Instances (stories/): 54 stories randomly generated with the skill, each set in a European city and published in three parallel versions — Chinese, English, and German — with illustrations for a subset of stories (the images/ case library). These stories are sample outputs demonstrating the skill's generative capability, not content built into the skill itself — the skill only contains the 48 experiment frameworks, and re-running it produces entirely new stories. Key fact: This is the first multilingual thought-experiment dataset in which the AI is the test subject. Every experiment specifies testable boundary conditions and behavioral principles that can be written directly into a system prompt — not vague philosophical musings.