What I learned asking 11 AI models to grade each other's AI predictions (shimin.io)

In an innovative experiment, a tech enthusiast asked 11 leading AI models to assess each other’s predictions on how AI will transform various industries. The models, including Claude Opus and GPT 5.4, tackled open-ended prompts that required them to analyze first through multiple order effects of AI implementation across sectors. Following this, each model was tasked with grading its peers on reasoning, originality, and correctness. Notably, Claude Opus 4.7 consistently emerged as the top performer, showcasing its advanced analytical capabilities and skeptical outlook on the speed of societal change brought by AI. This exercise not only underscores the varying strengths and weaknesses of advanced models but also introduces a qualitative layer to AI evaluation that extends beyond traditional benchmarks. The findings reveal that while top models excelled in self-critique, some, like Gemini 3.1, misjudged their performance—a phenomenon identified as the model delusion index. This analysis is significant as it prompts the AI/ML community to consider the implications of model self-awareness and its effects on AI's role in decision-making, encouraging a dialogue on the future of collaboration between humans and intelligent systems. For ongoing updates and a deeper exploration of model personalities, readers are directed to the AI-on-AI Arena.