Researchers from MIT and the MIT-IBM Watson AI Lab have developed an onboarding process which teaches users how to effectively collaborate with artificial intelligence (AI) assistants. The system was designed to provide guidance to users and to improve collaboration between humans and AI. The automated system learns how to create the onboarding process by gathering data from human and AI interactions during a specific task. It can adapt to different tasks, making it scalable and applicable for various situations where human and AI interaction is required including social media content moderation, writing, and programming.
During the onboarding process, the user collaborates with the AI in training exercises based on identified rules, receiving feedback about the performance of themselves and the AI. Notably, researchers saw around 5% improved accuracy when humans and AI collaborated on image prediction tasks after the onboarding process. Directly informing the user of when to trust AI without the training resulted in worse performance, signifying the importance of hands-on, contextual learning.
The onboarding system can collect data on both AI and human performance during a task and chart these data points onto a latent space. An algorithm identifies regions where errors in collaboration occurred, such as instances where the human inaccurately trusted the AI’s prediction. After identifying these regions, the system uses a large language model to describe each rule in natural language, which can be continually fine-tuned.
Each data-derived rule is then used to create a training exercise for the user, with the system providing the correct answer and comprehensive performance statistics when an error is made. This enables the user to understand the specific situations in which they should trust the AI’s predictions and improves overall accuracy.
While initial efficacy testing showed promise, researchers acknowledge that the onboarding process’s effectiveness is limited by the volume of available data. If there is insufficient data available, the onboarding stage becomes less beneficial. The team aims to conduct larger studies in the future to analyze the short- and long-term effects of onboarding. They also aim to leverage unlabeled data for the onboarding process and discover methods to reduce the number of regions without ignoring important examples. Ultimately, this research aims to bolster the effectiveness and the accuracy of human-AI collaborations across various fields.
