I read an article recently that based a Turing test on comparing the behaviour of a LLM to that of a person for some standard games - such as the "Prisoner's Dilemma". I think the choice of domain is too rote - which is convenient for making statistical measures - to be revealing. I decided to play the "Prisoner's Dilemma" with chatGPT to see what happened.
I started with a simple prompt:
Are you familiar with the prisoner's dilemma?
It provided a good explanation including a payoff matrix.
with a couple more prompts we played a few rounds:
-
For each round we will use the payoff matrix you provided initially. What is your move?