Scientific AI: Toward Recursive Epistemic Agents for Causal Discovery and General Intelligence

Artificial intelligence systems today are predominantly passive learners: they extract patterns from large datasets but lack the capacity for explanatory, causal, and counterfactual reasoning. In this paper, we argue that genuine understanding requires epistemic agency: the ability to form and revise hypotheses through active experimentation and model calibration. We introduce the framework of Scientific AI—agents that learn through discovery rather than observation—and propose a mathematically grounded architecture based on recursive hypothesis generation, causal inference, and multi-timescale calibration. We demonstrate this approach with a proof-of-concept symbolic physics environment, where the agent discovers novel laws through structured epistemic loops. Scientific AI provides a principled path to general intelligence rooted in explanation, not imitation.