Models made such astounding leaps in the past year in the areas of text and image generation that some believe that AI can now learn and create meaningful outputs independent of human intervention. It has even been claimed that models like LaMDA and ChatGPT demonstrate artificial general intelligence, or that they can fully replace the work of human designers or writers.

However, if you scratch the surface, such models still rely heavily on human intervention at every stage. In this talk, we’ll cover three tasks which cannot yet be automated since they rely on human judgment and expertise.

Firstly, we’ll go over how human intervention is needed when selecting and screening data to train these models, especially when it comes to spotting and removing bias or other quality issues. Secondly, we’ll talk about how people are needed to assess the ethical implications of models prior to their creation and deployment, from the consent to use the training data to whether certain models should be created at all. Finally, we’ll discuss how only human judgment can determine the most feasible uses of these models, and cover some real-world examples of where their application has been more, and less, successful.