HUMAN-AI INTERACTION OVERVIEW
The Ethics of Human-AI Interaction
If we designed AI systems that were morally perfect in a vacuum but didn't consider the predictable way people react when interacting and using those systems, then we would end up with very bad AI systems. Our conception of machine ethics and our design of interactive systems should aim to minimise harms and optimise goods in machines, ourselves and the outcomes of our interactions. This has two sides.
First, it's no good designing systems that would deliver optimal outcomes if they were used only by perfectly rational human operators. People have predictable cognitive habits and biases, as well as other prejudices and moral failures. If these aren't considered, then we will reliably end up aggravating harms and introducing injustices.
Second, we need to keep in mind the dynamic effects of human-AI interaction, most importantly the ways in which how we work with automated systems will change us. Our goal should be to design automated systems that help us act better, morally speaking, not to design systems that make better decisions than we would, on our behalf.
The first stage of research in this theme is identifying the human attributes that need to be considered when designing automated systems. There is, of course, a substantial field of work on human computer interaction, and 'human factors', which we will be engaging with—though our combination of philosophy, sociology, and social psychology will bring something new to that field. We're particularly interested in exploring the cognitive defects that AI systems must account for (and might exacerbate), as well as the human skills—including skilled moral behaviour—that automation might ultimately undermine.
The second phase is to determine what kind of interactions between humans and AI systems we want to aim for, what the goals and constraints are. For example, is there such a thing as moral skill, which can be weakened or strengthened by increased reliance on machines? In what other ways are our technologically mediated lives changing us, and how can we direct those changes for the better?
The design phase of this subproject will see us determine how we can shape machines so that they help make us better people, rather than just making better decisions on our behalf. It also means anticipating the kinds of cognitive errors that we are likely to make, as users of AI, and mitigating them in advance—for example by representing the actual degree of uncertainty in a choice more clearly, or by not just doing the right thing, but providing assurance and reassurance that the right thing is being done.
This research project brings together our established work on the sociology of affordances with new projects on role-taking and empathy in human-AI interaction, and on the ethics of using autonomous weapon systems, as well as on the prospects of moral deskilling and upskilling, and on the overlaps and synergies between artificial intelligence research and computational neuroscience.