Building Trust with a Teachable Artificial Intelligence: The Case of Repeated Trust Games
Description
This study explores the teaching of Artificial Intelligence (AI) systems in a repeated trust game. We evaluate whether participants trust the AI and teach it to adopt the most beneficial strategies among a set of four options with different levels of benefit. Results indicate that participants are initially cautious with the AI but increase their trust throughout the experiment, especially those with initially low trust. Participants imperfectly teach the AI, initially adopting beneficial strategies, then progressively learning the most advantageous ones while discarding those that offer minimal or no benefit. We also observe that participants inefficiently choose to avoid positive learning when it involves a risk of large losses. In additional tasks, we observe that participants naturally seek ways to exploit the AI, although a significant portion maintains human fairness in the interaction. Moreover, they effectively transfer prior knowledge to similar tasks. We conclude that participants trust an AI they can teach to generate significant benefits, although the risk of loss limits this trust and the efficiency of teaching.
Files
Steps to reproduce
The experiment was programmed with oTree (Chen et al. (2016)) and deployed online with Heroku. Participants were recruited through advertisements posted within local Telegram groups for research participants, and interested individ- uals were scheduled for participation at a mutually convenient date and time. We registered 321 participants for our experiment, with n = 253 participants participating and completing the experiment across three waves of data collec- tion. The first wave occurred from 30th August 2023 to 9th September 2023, the second wave occurred from 20th January 2024 to 25th January 2024, and the third wave occurred from 1st April 2024 to 17th April 2024. We conducted the experiment in multiple waves because it was part of a larger project that encountered a dwindling participant recruitment pool and we therefore needed to allow time for the participant pool to replenish. We implemented a pro- tocol to prevent individuals from entering multiple times and filtered out such occurrences. The experimental sessions were conducted online using Zoom. Par- ticipants were given the link to their experimental session upon confirmation of registration. The experimenter welcomed participants during the first five min- utes, read the experimental instructions, and created the session on Heroku. He then distributed each participant’s individual link to the experiment via private message in Zoom. Participants signed a consent form on the first screen and the experimenter monitored their progress throughout the experiment. Once par- ticipants reached the payment stage, they entered their personal information to receive the payment via Paylah, a local Singaporean app offering instantaneous transactions. The average experimental payment was 22.11 SGD, including a 5 SGD show-up fee. The payment was determined by randomly selecting one of the twenty periods from Task 1, Task 2, and Task 3. The points earned by par- ticipants in the selected period were then summed and converted into money at a rate of 1 point = 0.5 SGD. After receiving their payment, participants acknowledged its receipt and could leave the experiment.
Institutions
- Singapore University of Technology and Design