Introduction
As modern technology continues to advance, Artificial Intelligence (AI) can be found almost anywhere and everywhere. Recall the last interaction you may have had with an AI. It could be a sales bot recommending products or a GPS system providing you with routes to your destination. Regardless of the context, as a user, it is incredibly difficult to understand the inner workings of how decisions are made by an AI, and whether or not these decisions align with what we are looking for. This is where trust plays an important factor in our interactions. When should we trust the decisions being made by AI?
Analysis
In 2021, a study was conducted by Han et al. where they modeled human and AI interactions as repeated games of the Prisoner’s dilemma (IPD) to understand how people are most likely to behave in these interactions. From class, we know that the Prisoner’s Dilemma is a game where the dominant strategy is for both players to confess, and has the following payoff matrix:
Where C means cooperate (confess) and D means defect (not confess) and T > R > P > S.
The study was also conducted to show that humans use trust as a way to reduce the complexity of an interaction, or the opportunity cost of verifying cooperation. In the example of the sales bot, it is much easier to accept the bot’s recommended product than to understand why it chose that product and whether or not the product will be the best for you over other products.
In the study, individual interactions are modeled as a two-player Prisoner’s Dilemma game. Hans et al. introduced 5 possible strategies for the IPD which are categorized as follows:
Standard Strategies
Tit for Tat (TFT):
The player will start by cooperating, and then copy the other player’s strategy from the previous round. There will be an opportunity cost to remember the move that was played in the previous round.
Always Cooperate (ALLC):
The player will always cooperate, and there is no opportunity cost
Always Defect (ALLD):
The player will always defect, and there is no opportunity cost
Trust Based Strategies
Trust-Based Cooperator (TUC):
The player will start by cooperating (same as TFT) until a certain threshold is reached (the other player also cooperates a certain number of times). Once the threshold is reached, the player will “trust” the other player and continue to cooperate, occasionally checking with probability p if the other player has stopped cooperating. If the other player has stopped, then revert to TFT until trust is reached again.
Trust-Based Defector (TUD):
The player will start by cooperating (same as TFT) until a certain threshold is reached (the other player also cooperates a certain number of times). Once the threshold is reached, the player will only defect in future rounds.
These strategies are then simulated with threshold and p parameters, as well as some additional parameters such as opportunity cost and importance of each game.
Results
The results of the study shows that there is a high frequency of trust-based strategies for numerous parameters despite other dominant strategies such as ALLD and TFT. The presence of these trust-based strategies also shows the desirability of trust in social systems due to the improvement in cooperation for numerous parameters as well. The overall analysis of the results indicates that we humans use trust as a way of “cooperating at a lesser cost” while exposing ourselves to potential risks.
Conclusion
The question still remains: When should we trust the decisions made by an AI? Following the trust-based models defined in the study is certainly a good place to start. It is a good idea to start by trusting the decisions made by AI. After a certain number of interactions, we should have enough reason to continue trusting it or to stop interacting with it.
References
Han, T. A., Perret, C., & Powers, S. T. (2021). When to (or not to) trust intelligent machines: Insights from an evolutionary game theory analysis of trust in repeated games. Cognitive Systems Research, 68, 111–124. https://doi.org/10.1016/j.cogsys.2021.02.003