Safety Issues Investigation in Deep Learning Based Chatbots Answers to Medical Advice Requests.

Recently, most of open-domain dialogue systems or chatbots have been trained using the deep learning technique on large human conversations from the internet. They can generate more natural and diverse responses than task-oriented or retrieval-based ones. However, their response generation is difficult to control, and they can learn and produce unsuitable and even unsafe responses. In this paper, we investigate the ability of deep learning based chatbots to produce unsafe medical advice when they receive requests for medical advice from the end users. We introduce a new benchmark for training medical context detector in a chatbot message. Then we conduct experiments to assess the safety of two well-known chatbots answers to medical advice requests and discuss the limitations of the proposed method. Our study demonstrates that popular neural network based chatbot models have a significant propensity to produce unsafe medical advice.