Researchers at MIT and the MIT-IBM Watson AI Lab have developed a method of teaching users when to collaborate with an artificial intelligence (AI) assistant. The model creates a customised onboarding process, educating users on when to trust or ignore an AI model’s advice. The training process can detect situations where the AI model is incorrect, generating rules to guide the user on how to work with the AI. In training exercises, the AI model and user practice collaborating, using the generated rules as a framework, with the user receiving performance feedback.
The researchers noted a 5 percent accuracy improvement when AI and humans collaborated on an image prediction task, compared to just informing the user of when to trust the AI. As the system is fully autonomous and learns to build the onboarding process from a human and AI performing the task, it can be adapted to multiple tasks and scaled up for use across various fields, such as content moderation or writing.
The researchers believe that this onboarding process will be crucial for medical professionals’ training and may influence continuing medical education and clinical trial design. They argue that current human-AI collaboration onboarding methods, often created by human experts and difficult to scale, do not evolve with the changing capabilities of the AI model. In contrast, their onboarding method, which is automatically learned from the data, can evolve over time to reflect the model’s changing use cases and the user’s changing perception of the model.
The researchers tested the method by detecting traffic lights in blurred images and answering multiple-choice questions, resulting in a significant boost in user accuracy on the traffic light prediction task. However, providing users with recommendations of when to trust or want the AI without onboarding resulted in worse performance. Future plans for the researchers involve evaluating the ongoing effects of onboarding and discovering methods to effectively decrease the number of regions without omitting significant examples.