Bots Bargaining with Humans: Building AI Super-Bargainers with Algorithmic Anthropomorphization
研究提出一种生成对抗网络框架，训练机器人在谈判中追求更优经济结果的同时表现得像人类。实验发现，这种超级机器人比纯模仿或纯高效的机器人表现更好，但即使与人类无异，也可能引发“恐怖谷”效应，降低主观评价，影响长期客户关系。
As AI-powered negotiations spread, their psychological and relational impact remains unclear. The authors propose a novel generative adversarial network framework that trains a bot to aim for superior economic outcomes while appearing “human” (“algorithmic anthropomorphization”). In a bargaining game experiment, they compare this “superhuman” bot with two simpler alternatives: a bot that mimics human behavior and a purely efficient bot. The results show that (1) superficial anthropomorphization can make a bot seem human but does not improve subjective evaluations, (2) the efficient bot is so rational that it is easily exploited, undercutting its performance, and (3) the superhuman bot achieves superior economic results while appearing more human than actual humans. Yet even when bots act indistinguishably from humans, they may trigger an “uncanny valley” effect, lowering subjective evaluations regardless of performance. Because subjective evaluations predict future negotiation outcomes, these findings highlight the potential negative impact that AI bargaining algorithms can have on long-term customer relationships. The authors urge firms to measure more than objective outcomes when assessing AI negotiators.