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

## TL;DR

AI is nearing the ability to autonomously complete the scientific cycle, but human oversight remains crucial to ensure relevance and avoid biases.

## Contribution

The paper proposes graded autonomy in AI-led science to balance speed and human relevance while preventing epistemic collapse.

## Key findings

- AI can now explore complex hypothesis spaces beyond human intuition.
- Uncontrolled AI autonomy risks model and epistemic collapse due to recursive bias.
- Hybrid causal and neurosymbolic frameworks are needed to anchor AI to human priorities.

## Abstract

Artificial intelligence is approaching the point at which it can complete the scientific cycle, from hypothesis generation to experimental design and validation, within a closed loop that requires little human intervention. Yet, the loop is not fully autonomous: humans still curate data, set hyperparameters, adjudicate interpretability, and decide what counts as a satisfactory explanation. As models scale, they begin to explore regions of hypothesis and solution space that are inaccessible to human reasoning because they are too intricate or alien to our intuitions. Scientists may soon rely on AI strategies they do not fully understand, trusting goals and empirical payoffs rather than derivations. This prospect forces a choice about how much control to relinquish to accelerate discovery while keeping outputs human relevant. The answer cannot be a blanket policy to deploy LLMs or any single paradigm everywhere. It demands principled matching of methods to domains, hybrid causal and neurosymbolic scaffolds around generative models, and governance that preserves plurality and counters recursive bias. Otherwise, recursive training and uncritical reuse risk model collapse in AI and an epistemic collapse in science, as statistical inertia amplifies flaws and narrows the investigation. We argue for graded autonomy in AI-conducted science: systems that can close the loop at machine speed, while remaining anchored to human priorities, verifiable mechanisms, and domain-appropriate forms of understanding.