Should We Use Language Models In High-Stakes Decision-Making Scenarios?

Speaker: Max Lamparth
The emergence of generative language models, such as the one powering ChatGPT, has sparked widespread interest due to its potential implications for the future of work and society at large. The drive to automate decision-making is reaching high-stakes applications like military applications and mental health care, where non-zero error rates lead to individual failures with dire consequences and the potential to cause wide-scale harm. Thus, it is time to scrutinize and ask whether we should use language models in high-stakes decision-making scenarios. In this talk, I will dissect the question by studying how human decision-making differs from language models, if language models add their own dynamics to conflict situations, and whether they can recognize emergency/high-stakes user queries.