STAI CDT students Elfia Bezou Vrakatseli, Benedikt Brückner and Luke Thorburn presented a paper on their framework ‘SHAPE’ at the 20th European Conference on Multi-Agent Systems (EUMAS 2023). The paper is focused on the ethics of influence and provides a practical method for developers to think about ethical concerns when developing algorithms that potentially have influence.
They describe the focus of the paper, “SHAPE: A Framework for Evaluating the Ethicality of Influence” here, “Agents – whether they’re human or not – often influence other agents. While there are types of influence that are generally considered to be unethical such as blackmail, there are numerous scenarios in which it is less clear whether influence is morally or ethically acceptable. For example, it has been shown that large language models or recommender systems can easily exploit flaws in human decision-making to manipulate human agents, but whether such practices are acceptable is often notoriously difficult to decide. We make use of existing literature in the field of philosophy on influence and related topics to construct a framework for evaluating whether influence in a specific situation might be unethical.”
Their work also provides insights into methods for governing influence, which they hope will encourage people to think about ethical concerns when developing algorithmic systems that exert influence.
A highlight for Elfia, Benedikt and Luke was the opportunity to meet people from very different fields and to be exposed to research from other areas. They also enjoyed the setting of the conference in Naples, Italy, and said, “Naples itself is an amazing city, and the conference organisers made sure that everyone had a great time. The Royal Mineralogy Museum, where the conference took place, was pretty unique and, during the lunch breaks, the organisers also took us to other exhibitions in the building. The delicious lunch buffets and a conference dinner in the Naples sailing club left nothing to be desired.”
You can read their paper here: https://arxiv.org/abs/2309.04352