Beliefs about AI influence human-AI interaction and can be manipulated to increase perceived trustworthiness, empathy, and effectiveness
A repository for the paper "Beliefs about AI influence human-AI interaction and can be manipulated to increase perceived trustworthiness, empathy, and effectiveness," Nature Machine Intelligence 2023.
Author: Pat Pataranutaporn1, *, Ruby Liu1, 2, *, Ed Finn3, Pattie Maes1
1MIT Media Lab, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
2Harvard-MIT Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
3Center of Science and the Imagination, Arizona State University, Arizona, United States
* equal contribution, e-mail: patpat[at]mit.edu, rliu34[at]media.mit.edu
As conversational agents powered by large language models (LLMs) become more human-like, users are starting to view them as companions rather than mere assistants. Our study explores how changes to a person's mental model of an AI system affects their interaction with the system. Participants interacted with the same conversational AI, but were influenced by different priming statements regarding the AI's inner motives: caring, manipulative, or no motives. Here we show that those who imagined a caring motive for the AI perceived it as more trustworthy, empathetic, and better-performing, and that the effects of priming and initial mental models were stronger for a more sophisticated AI model. Our work also indicates a feedback loop where the user and AI reinforce the user’s mental model over a short time; further work should investigate long-term effects. The research highlights the importance of how AI systems are introduced can significantly affect the interaction and how the AI is experienced.
For more information: https://www.media.mit.edu/projects/beliefs-about-ai/overview/
Notebooks that have an Eliza and GPT version are essentially identical, and are merely copied for organization.
Processes the data from the CSV survey results and conversation transcripts and saves them into a new CSV.
Extract statistics about the Likert questions on the survey, including which statistical tests were used, p-value, mean, standard deviation for each item split by assigned motives and by perceived motives.
A few simple vizualizations of demographics.
Processes conversation data from the pre-processed data and saves them into a new CSV.
Vizualizes conversation data, generating sentiment trend line plots and box plots, in addition to calculating regression statistics. Note that some of the labels do not have a neat appearance; the labels of the final figures in the paper often were often remade in another application.
Calculate statistics and generate bar charts for Likert items. The code for calculating the statistics is in stat_process.py in the Include folder.
The notebook used for drafting the code for stat_process.py. Includes some potentially useful information for clarity.
Include/stat_process.py is used to calculate the statistics and generate plots, particularly for the Likert items of the survey.
requirements.txt contains the packages necessary to download.
chatlog.js a Google Apps Script to be run on Google Sheets to record data from a web interface. It can be added to Google Sheets from Extensions > Apps Script.
"Melu" is the chatbot interface used in the study. This folder includes the GPT3 and ELIZA versions of the Melu chatbot; each subfolder contains HTML, Javascript, and CSS (SCSS) code. This was run on CodePen to create a web interface, which you can see here (API codes redacted):
To use the GPT-3 CodePen, you will need to replace the YOUR OPENAI API KEY HERE with your OpenAI key. To collect the conversation data in a sheet, you will need to replace YOUR SHEET API KEY HERE with the URL generated from the Google Apps script.
Data is in the Results folder, split into Raw and Processed folders.
CSV data for surveys and conversations. The Survey data is from Qualtrics, and the Chatlog data is the transcript data from Google Sheets.
Data processed and saved from the 0_data_pre_processing notebooks, as well as conversation data processed and saved from the 3_convo_processing notebooks. []_Processed_Data.csv is the processed survey data, and []_Convo_Processed_Data.csv is the processed conversation data.
Note that conversation length labeled as "conversation turns" in the data and code, but this refers to the length.