The final workshop broadens the lens to the ethical and societal dimensions of deploying AI and machine learning in the physical sciences. While ethical AI is often discussed in the context of facial recognition, hiring algorithms, or social media, the physical sciences present their own distinctive challenges: dual-use risks in AI- accelerated materials discovery and synthesis planning, the environmental costs of large-scale model training, attribution and credit in human–AI collaborative research, and the risk of perpetuating biases embedded in historical datasets — for example, the overrepresentation of certain element combinations or synthesis routes in public materials databases.

The session addresses bias in training data, examining what enters public materials databases, what is excluded, and how these selection effects propagate into model predictions and discovery recommendations. Dual-use risks are considered in the context of ML-guided synthesis planning and materials design, where the same tools that accelerate beneficial discovery can in principle be directed toward harmful applications. The environmental cost of training large ML models is quantified and contextualised. The workshop also examines the changing nature of scientific authorship and peer review in the age of AI-generated content, addressing questions of attribution, transparency, and the integrity of the scientific literature.

Relevant regulatory frameworks, including the EU AI Act and its implications for research institutions, are discussed alongside principles for responsible deployment: uncertainty disclosure, human oversight, and re- versibility. Participants will gain a structured framework for identifying and managing the ethical dimensions of ML deployment in their own research programmes.

Extended Key Takeaways & Outputs

• Understanding of bias sources and propagation pathways in scientific ML datasets

• Assessment of dual-use risks in AI-accelerated materials discovery and synthesis planning

• Quantification of the environmental cost of large-scale model training

• Guidance on AI authorship, attribution, and scientific integrity

• Awareness of the EU AI Act and relevant regulatory frameworks for research institutions

• Principles for responsible deployment: uncertainty disclosure, human oversight, and reversibility