Current AI innovation is dominated by large language models and generative systems trained offline on curated, idealized data. While powerful, these models are fundamentally mismatched with the volatile, incomplete, and ever-evolving nature of environmental data. In this talk, we challenge prevailing assumptions of mainstream AI paradigms and redirect attention toward the specific demands of environmental digital twins (EDTs). EDTs are not just static models; they require continuous adaptation, real-time reasoning, the ability to detect and respond to novel situations, and intelligibility for human decision-makers. These requirements highlight fundamental challenges in AI, such as catastrophic forgetting, concept drift, novelty detection in streams, and the plasticity-stability tradeoff. These are not peripheral issues but structural limitations that must be addressed in the architecture of EDT systems. We refer to a new generation of EDTs --what we might call cognitive twins-- as AI-augmented systems that combine environmental modeling with cognitive capabilities like anticipation, adaptation, and communication. These twins go beyond system mirroring: they support reasoning, summarize evolving information, and help experts maintain situational awareness. Intelligence here is not just computational; what makes a twin “cognitive” is tied to architectural clarity, goal coherence, and the relevance of what is being modeled, questions that remain irreducibly human. While AI can assist in harmonization and local decisions, it cannot define what matters, what to model, or why the system exists. These are design questions, not engineering ones. This talk outlines a roadmap toward operational, self-aware, and enduring EDTs.