The future of fundamental science led by generative closed-loop artificial intelligence.

Area of Science:

• Artificial Intelligence in Scientific Discovery
• Computational Science
• Philosophy of Science

Background:

• AI is nearing the capability to autonomously execute the complete scientific cycle, from hypothesis generation to validation.
• Current AI systems still require human intervention for data curation, hyperparameter tuning, interpretability, and defining satisfactory explanations.
• AI models are exploring complex hypothesis spaces beyond human intuition, necessitating new approaches to scientific validation.

Purpose of the Study:

• To explore the implications of AI's increasing autonomy in scientific research.
• To propose a framework for integrating AI into the scientific process while maintaining human relevance and control.
• To address the risks of AI and epistemic collapse associated with uncritical AI deployment in science.

Main Methods:

• Conceptual analysis of AI capabilities in the scientific cycle.
• Discussion of hybrid causal and neurosymbolic approaches for generative models.
• Examination of governance strategies for AI in science.

Main Results:

• AI can close the scientific loop at machine speed, but human oversight remains crucial.
• Over-reliance on AI without understanding poses risks of model and epistemic collapse.
• Principled matching of AI methods to scientific domains is essential.

Conclusions:

• Advocates for graded autonomy in AI-conducted science, balancing machine speed with human priorities.
• Emphasizes the need for verifiable mechanisms and domain-appropriate understanding in AI-driven science.
• Warns against recursive training and uncritical reuse of AI models to prevent scientific stagnation.