When communicative AIs are cooperative actors: a prisoner’s dilemma experiment on human–communicative artificial intelligence cooperation

ABSTRACT

This study examined the possibility of cooperation between human and communicative artificial intelligence (AI) by conducting a prisoner’s dilemma experiment. A 2 (AI vs human partner) × 2 (cooperative vs non-cooperative partner) between-subjects six-trial prisoner’s dilemma experiment was employed. Participants played the strategy game with a cooperative AI, non-cooperative AI, cooperative human, and non-cooperative human partner. Results showed that when partners (both communicative AI and human partners) proposed cooperation on the first trial, 80% to 90% of the participants also cooperated. More than 75% kept the promise and decided to cooperate. About 60% to 80% proposed, committed, and decided to cooperate when their partner proposed and kept the commitment to cooperate across trials, no matter whether the partner was a cooperative human or communicative AI. Overall, participants were more likely to commit and cooperate with cooperative AI partners than with non-cooperative AI and human partners.

KEYWORDS:

• Artificial intelligence
• computers are social actors
• cooperation
• human–AI interaction
• human–machine communication
• social dilemmas

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 In prisoner’s dilemma experiments the words “cooperate” and “defect” are often avoided and are replaced by undetermined characters or colours as they bias a player. However, some previous prisoner’s dilemma experiments on human–machine interactions (e.g., Miwa and Terai 2012; Sandoval et al. 2016) adopted the approach the current study used and their results did not bias towards cooperation because of the framing used.

2 It is acknowledged that it was not necessary to pay the participants all the max payment. The action may influence the performance of participants who will participate in similar studies in the future if they foresee similar payment.

Additional information

Funding

This study was supported by the Tier 2 Start-up Grant of Hong Kong Baptist University [grant number RC-SGT2/19-20/COMM/001].