Massachusetts Institute of Technology (MIT) researchers explored humans' ability to play the cooperative card game Hanabi with an advanced artificial intelligence (AI) model trained to play with teammates it has never encountered before.
Single-blind experiments matched humans with either a reinforcement learning-based AI agent as their teammate, or a rule-based agent; neither match-up scored better than the other.
People consistently disliked playing with an AI teammate, as strange or seemingly illogical moves by the AI undermined humans' trust in it.
Such animosity could spell trouble for researchers designing technology to collaborate with humans on real challenges.
MIT's Ross Allen said if researchers do not address the issue of subjective human preference, "then we won't create AI that humans actually want to use."
From MIT News
View Full Article
Abstracts Copyright © 2021 SmithBucklin, Washington, DC, USA