AI Research Institute on Interaction for AI Assistants (ARIA)

Summary

Establishment of the AI Research Institute on Interaction for AI Assistants (ARIA) to develop next-generation AI assistants for mental and behavioral health, focusing on trust, empathy, and personalization.

What they want

ARIA will integrate human and machine cognition studies to advance AI assistants for mental and behavioral health. Research activities are centered around three pillars: Grounding (developing new models for efficient learning, causal models, computational theories for trade-offs, and evaluation metrics for trustworthy AI), Instructability (designing new paradigms for establishing trust in AI, theories of human-AI interaction, methods for describing AI's internal processing, and models/training procedures for AI assistants), and Alignment (advancing human-centered design, establishing precise definitions for alignment, developing computational/experimental methods to operationalize definitions, and creating metrics of alignment in ethical/social contexts). The institute will also grow a future-ready workforce through interdisciplinary education pathways from K-12 through postgraduate training.

Deliverables

New models for efficient learning and generalization
New learning algorithms leading to rich, causal models
Computational theories for navigating trade-offs between learning algorithms or model architectures
New evaluation metrics for tracking progress toward trustworthy AI assistants
New paradigms centered on establishing trust in AI
New theories and models of how humans interact with AI
New methods for describing AI’s internal processing
New models and training procedures for integration into a computational framework for AI assistants
Advanced current best practices for human-centered design
Precise definitions for what it means to be aligned
Computational and experimental methods to operationalize alignment definitions
Cognitively and computationally sound metrics of alignment in complex ethical and social contexts

Technical requirements

New models for efficient learning and generalization
New learning algorithms leading to rich, causal models
Computational theories for navigating inherent trade-offs between learning algorithms or model architectures
New evaluation metrics for tracking progress toward trustworthy AI assistants
New paradigms centered on establishing trust in AI
New theories and models of how humans interact with AI
New methods for describing AI’s internal processing
New models and training procedures for integration into a computational framework for the development of AI assistants
Advancement of current best practices for human-centered design
Establishment of precise definitions for what it means to be aligned

Key personnel

Researchers in computer science
Researchers in neuroscience
Researchers in cognitive science
Researchers in philosophy
Researchers in law
Researchers in education
Mental health practitioners
Civil society groups