Artificial intelligence models have the capability to discern patterns in images more effectively than the human eye in many cases, although there are exceptions. When a radiologist utilizes an AI model to assist in evaluating whether a patient’s X-rays exhibit indications of pneumonia, it becomes crucial to determine when to rely on the model’s recommendations and when to disregard them.
Researchers at MIT and the MIT-IBM Watson AI Lab propose that implementing a personalized onboarding process can aid the radiologist in making this determination. This system is designed to educate the user on when to engage with an AI assistant effectively.
Through this training approach, scenarios may arise where the radiologist places trust in the model’s advice, even when it is inaccurate. The system autonomously learns the guidelines for collaborative interaction with the AI and articulates them using natural language.
During the onboarding phase, the radiologist engages in collaborative exercises with the AI based on these guidelines, receiving feedback on her performance and that of the AI.
The study conducted by the researchers demonstrated that this onboarding strategy resulted in approximately a 5 percent enhancement in accuracy when humans and AI collaborated on image prediction tasks. Conversely, merely informing the user when to trust the AI, without adequate training, led to inferior performance.
The automated nature of the researchers’ system enables it to adapt the onboarding process based on data derived from the human-AI task performance. This adaptability allows for scalability and application in various human-AI collaborative scenarios, such as social media content moderation, writing, and programming.
The researchers emphasize the significance of incorporating such onboarding procedures into the training regimen of medical professionals. They suggest that doctors utilizing AI to make treatment decisions may need to undergo similar training to ensure optimal collaboration with the technology.
The evolving nature of the training method distinguishes it from existing onboarding techniques for human-AI collaboration, which often rely on static training materials created by human experts for specific use cases. By automatically learning from data, this method evolves over time to accommodate changes in the AI model’s capabilities and the user’s perception.
Moving forward, the researchers aim to conduct extensive studies to assess the immediate and long-term impacts of onboarding. They also seek to leverage unlabeled data for the onboarding process and develop strategies to streamline the identification of regions without overlooking critical examples.
