Biologically-Inspired AI Safety
Summary
Biologically-Inspired AI Safety is an approach that draws insights from biological organisms and neuropsychology to develop safer and more aligned AI systems. This research agenda combines bottom-up and top-down approaches. The bottom-up component involves creating biophysically realistic simulations of simple organisms like nematodes, fruit flies, and zebrafish to inform AI algorithms and system architectures. The top-down component focuses on grounding AI goal structures in neuropsychology to improve value alignment. By pursuing these parallel tracks, researchers aim to develop AI systems that are inspired by embodied organisms with sensorimotor integration, potentially leading to more robust and aligned AI. This approach has the added benefit of being rooted in established research traditions and aligning with current themes in biological and psychological sciences, such as data/model integration and reproducibility.