AI is a gamechanger in the quest for better understanding Earth data. ML allows training of models for virtually any purpose, and many of them are published on open platforms like HuggingFace and Kaggle. However, in practice it is not easy particularly for non-experts to use such models, due to several blockers: Models typically need highly specific data preprocessing requiring skillful python coding. Model metadata are sparse and not standardized. In particular, they are not machine-readable so human intervention is required. A model's comfort zone is not always delineated clearly, and outside of it model accuracy and reliability can drop drastically, such as below 20%.

Recent work in research and standardization is aiming at overcoming these obstacles in the quest for easy-to-use, zero-coding, reliable ML use on spatio-temporal Earth Data. Based on ongoing research in the EU-funded FAIRgeo project we discuss AI-Cubes as a novel paradigm which embeds ML inference seamlessly into the geo datacube query standard, WCPS. Further, the concept of Model Fencing aims at deriving hints about a model's comfort zone so that the server can automatically decide about model applicability on the region selected and warn the user.

Live demos, several of which can be recapitulated by the audience, serve to illustrate the challenges and solution approaches. Ample time will be reserved for active discussion with the audience.