SUMMARY

Our panel will address the question of “when will we trust AI?”. We will do this by teasing out what makes a given AI tool worthy of trust, in collaboration with the audience. We will particularly focus on the fact that “AI” is not a monolith, but rather a collection of different technologies and tools with very different designs, capabilities, and risks that vary by use and context.

With the help of a polling tool, we will pose questions to the audience designed to draw out implicit assumptions and intuitions about AI and trust, including inviting audience members to share their views. In conversation with the audience, our panel members will clear up common misunderstandings about the specific risks, limitations, and failure modes of different AI technologies.

For example, we will discuss the differences in how to assess the trustworthiness of AI models that are:

1) designed to mirror changing human preferences,

2) designed to optimize for a prespecified measure of task success,

3) designed to generate novel variations on learned patterns. We will also elicit a discussion of the ways each kind of model could potentially be “trustworthy”.

Through active participation, audience members will leave with a better grasp of the diverse types and applications of AI and an understanding of what questions they should be asking to assess the trustworthiness of different AI tools and use cases in their own lives.