ITB
For students and young adults (15–25) who read AI-written text every day in chats, comments and group work, and for the teachers and AI-literacy trainers who have to teach them to spot it. Today they have no way to practise or measure that skill: AI-literacy sessions are passive lectures and quizzes, and most people overestimate how well they detect AI text. Existing social-deduction games (Among Us etc.) only have human impostors, so they never test this. Nothing gives a group a live, playful setting where AI hides among real players at adjustable difficulty, and nothing records how often the AI was caught per difficulty level.
Preliminary score
- Value for the user16.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI23.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing13.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility9.0 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- GPT-6 AstraPreliminary71/100
- Strengths
- Your structured Q&A meetings make AI behaviour central to the game rather than an optional chatbot feature. Your documented 148 checks target concrete risks, including role leakage, duplicate submissions, provider failures and malformed network traffic. Your difficulty levels, persona memory and per-round detection logging provide a credible foundation for balancing the experience.
- To improve
- Provide a complete match trace showing live model calls, voting and the corresponding analytics. Use your logs in human playtests to report detection and false-accusation rates by level and language, then compare learning outcomes with a quiz-based baseline. Validate that improvements transfer to unseen AI text rather than recognition of your injected typos, timing heuristics or repeated fallback phrases. Report per-match inference and hosting costs, response latency and behaviour under concurrent-room load.
- Claude Opus 5.5Preliminary66/100
- Strengths
- Your core loop is playable: the screenshots show a deployed game with lobby, tasks, Q&A meetings and voting, and AI is the impostor itself rather than decoration. The AI design shows real thought: per-level humanised timing, typo and lowercase injection, persona memory, a self-reveal filter with fallbacks, and no AI flag on the wire. Your 148 automated checks in CI cover failure paths such as LLM throw, hang, empty output and self-reveal, plus protocol abuse, which is rare at a hackathon. The JSONL analytics with per-level suspectedRate and votedOutRate give you the right tool to measure whether the disguise works.
- To improve
- Your testing is strong on engineering but says nothing about the game's actual purpose. Run playtests and report real detection rates per AI level, the questions or answer patterns that gave the AI away, and failures you found. Compare the game with a standard AI-literacy quiz or lecture, even with a small group, to back the claim that it teaches people to spot AI text. Add running costs per match (Gemini and Groq calls per meeting and per movement tick) and name a concrete next step, such as a pilot with one school.
The team
Aykhan Nazaraliyev
developer
Ashraf AbbasovLead
business
Kenan Salmanli
data
Farid Gahramanov
developer
Shikhi İbrahimov
product