The Unpredictability of Large Language Models: A Double-Edged Sword for Service Delivery


After receiving an order to pay compensation for misleading a customer with inaccurate information from their chatbot, Canada's largest airline faces a challenge in utilizing generative AI in real-world situations. While it is seen as a helpful assistant by many, the potential risks prevent it from fully taking over services. The UK Government Digital Service conducted an experiment with a large online model and found that 70% of users found the chatbot's results useful, but this does not guarantee reliability.

This highlights the difficulty in controlling the output of chatgpt even after one year of widespread adoption. Some argue that because it is a stochastic process, it is logical to have unpredictable outcomes, while others find it intriguing that chatgpt remains a blackbox even for experts. This has led to various funny experiments, including providing tips to improve chatbot results. The machine seems to care about the tip, even if it doesn't benefit from it.

Image source