Agonistic AI: Advancing Interpretive Pluralism in the Cultural AI Value Space

AI sycophancy, the structural tendency of generative AI to excessively validate users' claims, poses a major concern for the future of cultural AI. Yet this phenomenon reveals that AI models encode a distinct normative position within their outputs, one that can be redirected into a heuristic for training and learning from human-AI (hAI) interactions as repositories of cultural knowledge. Here we propose understanding hAI interactions as agonistic--inherently contested and value-pluralist--against conventional consensus-driven models. Against the "reward model" of reinforcement learning, we extend the adversarial logic already present in AI training to the construction of hAI interactions as an evolving and contested value space. To do so we draw on humanistic methodologies (cultural analytics) as a technology for abstracting and evaluating normative content from AI output. We show how these methods can be integrated into AI system design in order to advance interpretive pluralism in an increasingly AI-mediated world.