Syntax, Not Science: Safeguarding Scientific Integrity via the Scientist-AI-Loop

par Nico SCHUSTER

lundi 18 mai 2026, 14:00 → 16:00 Europe/Paris

Abstract:
Integrating Large Language Models (LLMs) into modern research workflows presents a critical challenge. While AI agents excel at functional syntax, they lack scientific intuition, frequently hallucinating scientific logic or silently discarding governing physical laws to optimize performance. To safely harness LLMs without producing invalid tools, we advocate for the Scientist-AI-Loop (SAIL). This human-in-the-loop framework structurally decouples scientific logic from coding syntax. The researcher acts as the conceptual architect, enforcing theoretical boundaries and identifying breakdowns, while the AI handles implementation. Validated via two visualization tools, a real-time gravitational lensing application and a dynamic cosmic structure formation simulation, SAIL provides a domain-agnostic blueprint broadly applicable to science, from particle physics to cosmology and other disciplines. During development, SAIL exposed critical, invisible AI failures where agents confidently fabricated physics. By structuring the progression from initial concept to final code, SAIL compresses development timelines from months to mere days. More importantly, it safeguards scientific integrity, ensuring that generative tools can finally be trusted for professional modeling and science communication.

Test the tools here:

• Gravitational Lensing Application: https://nicosmo.github.io/lensing_visualization/
• Cosmic Web Explorer: https://nicosmo.github.io/cosmic_web_explorer/