Researchers at MIT and the MIT-IBM Watson AI Lab have developed a system that educates a user on when to trust an AI assistant’s recommendations. During the onboarding process, the user practices collaborating with the AI using training exercises and receives feedback on their and the AI’s performance. This system led to a 5% improvement in accuracy on an image prediction task when humans and AI collaborated. The process is fully automated and tailored to the data from the human and AI performing specific tasks, making it adaptable to various situations. This system could be used in numerous fields where humans and AI work together, such as social media content moderation, writing, and programming.
Graduate student and lead author Hussein Mozannar explains that currently, many people are given AI tools to use without any training on when they can be beneficial. The researchers aim to address this lack of guidance from both a methodological and behavioral perspective. The team believes that their onboarding process will become an integral part of training for professionals, particularly in healthcare. They suggest that doctors collaborating with AI on treatment decisions may first need to undergo training similar to their onboarding process.
Existing methods of onboarding for human-AI collaboration often involve training materials created by human experts for specific use cases, which can be challenging to scale. The researchers’ onboarding method, in contrast, is automatically learned from data and evolves over time. It begins by collecting data on the human and AI performing a specific task, then uses an algorithm to identify instances where the human incorrectly collaborates with the AI. These instances are described as rules in natural language which are then used to build training exercises. This process significantly improved users’ accuracy on tasks, but giving recommendations without onboarding led to worse performance and confusion among users, possibly due to resistance to being told what to do.
In the future, the team plans to conduct larger studies to assess the short- and long-term effects of onboarding. They also want to utilize unlabeled data for the onboarding process, and find methods to reduce the number of regions effectively without omitting important examples. This research is partially funded by the MIT-IBM Watson AI Lab.
