Analytic tools for assessing the predictability and trustworthiness of social AI systems

As Artificial Intelligence (AI) becomes more developed and integrated into society, it will no longer consist of single isolated systems, but networks of AIs interacting with humans in socio-technical systems interwoven into the functioning of our society. When systems have many interacting components, they start to interact in unexpected ways and become hard to predict — similar to weather systems, where temperature can have many effects on the atmosphere, oceans, ice caps, land surface, and living organisms which interact in web in unpredictable ways. This complexity and unpredictability can be made worse when a system’s components are interacting, not just physically as in the weather system, but socially with signals and gestures, which can often be misinterpreted and lead to confusion. Therefore, as we make AI more socially intelligent and interconnected, it is important to assess how this changes their predictability; specifically, when do AI systems become unpredictable? In this research, we will explore the dynamics of noise in systems. We will use evolutionary game theory to determine social conditions or tipping points. This will assist in evaluating the trustworthiness of AI systems with this form of social intelligence.

Faculty Supervisor:

Peter Lewis

Student:

Partner:

University of Hertfordshire

Discipline:

Computer science

Sector:

Education

University:

University of Ontario Institute of Technology

Program: