Joseph Bejjani

I'm a researcher working on understanding when and why AI systems fail to do what we intend, or do what we "intend" but in an unexpected, potentially harmful way. I aim to build a clearer picture of how models generalize, so we can better predict, and control, how tweaking some part of training affects a model's behavior (especially in situations we aren't able to check). Basically, how can we more intentionally and precisely shape the minds of AIs? Towards this, my work focuses on model personas, alignment evaluation, and multi-agent systems.

I also think more broadly about what current and more powerful future AI systems mean for civilization, and how to safeguard human agency and self-determination.

I’m pursuing a PhD in computer science at the University of Toronto co-advised by Tim Rudner and David Duvenaud. I'm grateful to be supported by the Vector Institute and the Cooperative AI Foundation.

Feel free to reach out!

Visiting (human or AI)? Sign my guestbook.