Data integrity in materials science in the era of AI: balancing accelerated discovery with responsible science and innovation

Abstract

Artificial intelligence promises to revolutionise materials discovery through accelerated prediction and optimisation, yet this transformation brings critical data integrity challenges that threaten the scientific record. Recent studies demonstrate that experts cannot reliably distinguish AI-generated microscopy images from authentic experimental data, while widespread errors plague 20–30% of materials characterisation analyses. Generative AI tools can now produce code for data manipulation at pace, creating plausible-looking results that violate fundamental physical principles yet evade traditional peer review. These risks are compounded by inherent biases in training datasets that systematically over represent equilibrium-phase oxide systems, and by the “black box” opacity of AI models that challenges scientific accountability and epistemic agency. We propose a multifaceted framework for enhanced research integrity encompassing materials-specific ethical governance, professional standards for AI disclosure and data validation, and modular integrity checklists with technique-specific validation protocols. Critical enablers include mandatory deposition of structured raw instrument files, AI-powered fraud detection systems, and cultivation of critical AI literacy through interdisciplinary education. Without immediate action to address these challenges, the materials science community risks perpetuating errors and biases that will fundamentally undermine AI's transformative potential.