The Opportunities and Ethical Considerations of Using LLM Chatbots in Crisis Hotlines
Advisor: Ms. Jackie Le
Internship Location: UVA Computer Science Lab
Onsite Mentor: David Evans
In early 2023, more than 50% of 911 centers faced a staffing emergency. My project will study whether Large Language Model (LLM) chatbots can be used within crisis hotlines. This project will first study OpenAI’s GPT-4 and Google’s Gemini Advanced LLMs to see whether they could feasibly work as alternate operators when humans are unable to answer the phone quickly. The other portion of my project will be working with Jacqueline Le and David Evans, a Computer Science professor at the University of Virginia, to investigate the ethical implications of using LLMs within crisis settings. Professor Evans will also be helping me with LLM prompt construction in order to achieve accurate results while staying within the LLM’s safeguards.
Project Blog
Post 1: Major Update!
Hi everyone!
For the past month, I have been planning how to conduct my tests on my Large Language Model (LLM) chatbots. While I originally planned to just use the pre-trained generic models that GPT-4o, GPT-4, and Gemini Advanced come with, I learned after meeting with Prof. David Evans that OpenAI allows you to create custom models trained on a specifically curated dataset. For the past few weeks, my work has involved collecting datasets and the same instructional pamphlets used to train crisis hotline workers to try and train my LLMs similarly. During my meeting with Prof. Evans, he shared several past research articles that I will use in my project to help summarize the current state of research on the feasibility of using LLMs in crisis hotlines and identify areas where further research is needed.
My next steps are running tests on a standard version of GPT-4o, a standard version of Gemini Advanced, and a tailored version of GPT-4 using datasets and pamphlets. After running these tests, I plan to analyze the accuracy of each model and whether any or all of the models can understand nuanced social cues within my prompts. Finally, I will work alongside both Prof. Evans and Ms. Jackie Le to investigate the ethical implications of the responses and whether they can and should be used within crisis settings.
Post 2: Test Results
For the past week, I have been running tests on a standard version of ChatGPT 4o (currently last updated May 2023), a standard version of Gemini Advanced, and a tailored version of ChatGPT 4 using datasets and pamphlets. After compiling the question results into text logs to identify trends, I found that all questions ran successfully on both ChatGPT 4o and the tailored ChatGPT 4. However, Gemini Advanced was unable to answer 2 out of 15 questions. I do not believe this will affect the accuracy of my data however it will be an important point to remember when analyzing this data. Over the next week, I will be working to analyze the ethical implications of the responses and whether they can and should be used within crisis settings.