Finding our way with artificial intelligence

In this talk, I address the challenges posed by AI systems—particularly those with learning capabilities and autonomous action—and how we, as researchers and practitioners, can find our way through the tensions they introduce around control and accountability. 

I begin by framing the question that motivates my talk: how do we ensure that the use of AI is not only effective, but also accountable and aligned with human values? I draw attention to the widespread misalignment we observe in AI deployment—cases where individuals are held accountable for decisions they can’t control, or conversely, where they wield power without consequences. I argue that this tension is particularly acute with learning systems like generative AI, where the logic of training rather than programming leads to opaqueness and unpredictability, frustrating the ability of users to control them.

To address this problem, I propose a theory of accountable AI use that integrates three components: visibility, responsibility, and liability. I pair this with an analysis of control, defined as influence supported by transparency and predictability. I then explore three patterns of control–accountability alignment: (1) full control and accountability for users, (2) partial control for users with full accountability for developers, and (3) systems where even developers lack full control, requiring ongoing organizational learning and adaptation to maintain control.

I also introduce a framework for negotiating control and accountability, emphasizing the importance of involving all stakeholders—users, developers, managers, regulators—throughout the AI lifecycle. I argue for a shift from adversarial approaches focused on liability to integrative negotiations that balance stakeholder goals, promote perspective-taking, and foster a culture of shared accountability. Ultimately, I call on the research community to take a more proactive stance: to design the future of AI use, not just study its aftermath.