Combat medics and AI alignment
Research
Project Duration
24 Months
Services Provided
Research into the best way to understand and represent individual differences to an AI for the purposes of alignment.
Problem
With the rise of AI, there is increasing concern over what has been called "AI alignment," which is the difficulty of getting an AI to fully understand and embrace human values. In an effort to find ways to better tackle this issue, DARPA created the "In The Moment" program with the goal of teaching an AI to embody the traits of an individual so that it can better replicate what that individual would do in an out of sample situation. This ongoing work is very theoretical and early stage. There are no plans to implement this anytime in the near future.
Solution
With the ShadowBox team, we conducted ~20 Cognitive Task Analysis (CTA) interviews with medics to drill down into the most difficult triage situations they had been involved in either as a civilian or combat medic. We used these interviews to discover individual differences in the context of triage, as well as to use them as building blocks to create our own scenarios. Having built those scenarios, we interviews a further ~20 medics and had them go through these custom built scenarios to further understand and validate differences in decision-making. The ongoing effort has identified major differences in medical sensemaking and decision-making, and we are working closely with other organizations to identify the best ways to represent these differences to an LLM in a high fidelity way. We are currently developing a survey to understand individual differences in sense-making to better understand how differences in what individuals identify as relevant could impact how they frame trade-offs.