Guided intelligence

The future of safe AI relies upon models that:

• behaves according to intention (specification)
• can withstand perturbations (robust)
• whose activity can be monitored and analysed (assurance)

Space of safety problems (Ortega et al 2019)