Why do you care about AI Existential Safety?

My research is about designing AI agents to achieve socially desirable objectives. As such, I want to optimise interaction to satisfy objectives that take into account the welfare of everyone involve. If AI is not safe and can be for example manipulated to the benefit of some stakeholders at the expenses of others, this would not satisfy important fairness properties. As such AI safety is a necessary condition for the systems I have in mind.

Please give at least one example of your research interests related to AI existential safety:

We often find ourselves in a position to choose between the ethical option – when we contribute to a common good by paying some cost, for example recycling – and the selfish option – when we instead do what is convenient to us free-riding on the work of others, for example not recycling.We know peer pressure plays a role in our decisions. But can we find out when it plays a positive role? With Jack Bara and Giulia Andrighetto, we found that some social networks promote cooperation better than others. Especially when we are allowed to choose who to interact with, and we can do so frequently enough. The good behaviour is quickly suppressed in a Barabasi-Albert Graph but it’s significantly more resistant in a closed community, under identical partner selection and imitation rules. Our computational framework to study the emergence of pro-social behaviour through imitation when individuals interact in a social networks was published in JAAMAS’23 and selected for the special JAAMAS track at AAMAS’24 (Accessible from https://dl.acm.org/doi/abs/10.5555/3545946.3598899).