MISID: A Multimodal Multi-turn Dataset for Complex Intent Recognition in Strategic Deception Games

MISID addresses a high-value niche in AI safety and behavioral analysis: Theory of Mind (ToM) and deceptive alignment. While sentiment analysis and basic intent recognition are commoditized, 'strategic deception' in multi-turn multimodal contexts (likely inspired by games like Among Us or Werewolf) is significantly harder to model. The project has 0 stars but 6 forks only 3 days after release, which is a classic signal of a research team or early academic reviewers prepping for collaboration. The defensibility score of 4 reflects that while the dataset itself is a proprietary/curated asset, it lacks a surrounding software ecosystem or 'data gravity' that would prevent a larger lab from generating a 10x larger synthetic or human-annotated version. The moat here is the difficulty of high-quality annotation for deceptive intent, which is labor-intensive. Frontier risk is medium because while OpenAI/Anthropic are obsessed with 'deceptive alignment,' they tend to focus on model behavior rather than human game datasets. However, as multimodal models (GPT-4o, Gemini 1.5) improve, the need for specialized datasets like this might be eclipsed by the models' inherent zero-shot reasoning capabilities in strategic contexts.