AAAI Publications, 2018 AAAI Spring Symposium Series

Font Size: 
Architecting a Human-Like Emotion-Driven Conscious Moral Mind for Value Alignment and AGI Safety
Mark R. Waser, David J. Kelley

Last modified: 2018-03-15


A general intelligence possesses the abilities, given any goals and environment, to iteratively evaluate, plan, discover or learn and build or gain competencies, tools and resources to succeed at those goals. The only known examples of general intelligence are the obligatorily gregarious, conscious “selves” designated homo sapiens that currently dominate our planet. We argue that humans are reasonably deep in a safe and effective attractor in the state space of intelligence and that adhering as closely as possible to the human model of an emotion-driven conscious moral mind, has the advantages of safety, effectiveness, comfort and ease of transition due to a known and explored state space. Most concerns about AI safety are due to expected differences from humans – which seems unnecessary when, not only can we choose to make them more humanlike but the history of AI research clearly shows that we are unlikely to succeed unless we do so. We therefore propose a human-like emotion-driven consciousness-based architecture to solve these problems. We rely upon the Attention Schema Theory of consciousness and the social psychologists’ functional definition of morality to create entities that are reliably safe, stable, self-correcting and sensitive to current human intuitions, emotions and desires.


AGI Safety; AI Safety; Value Alignment; Moral Machines; Emotion; Consciousness; HUman-like

Full Text: PDF