Babies currently have more common sense than artificial intelligence

Results of an experiment performed by researchers at the New York Univeristy show that even 11-month old children currently have more common sense than artificial intelligence systems. Pim Haselager, professor of AI at Radboud Univeristy, features in De Nieuws BV on NPO Radio 1 to talk about these findings.

Haselager explains how the experiments were performed based on what expectations 11-month olds held, which were measured by the time they spent looking at an event. In this case, clips were shown of objects taking detours to get to their destination due to the existence of obstacles. However, when the obstacles disappeared and the object would still take a detour to its destination, this surprised the babies. This showed that babies have grasped the efficiency principle.

Aside from this, neural networks were shown the same clips. The AI algorithms were made to learn and predict what path the object would take with and without the obstacles present. The results showed that the algorithms were unable to predict a shorter path in the absence of any obstacles, showing that they do not yet grasp the same efficiency principle. Understanding that detours are unnecessary in the absence of obstacles is part of human common sense, which appears to be specifically hard to model in AI.

Haselager further talks about why it seems models like ChatGPT appear to have more common sense and even emotion, without actually understanding what is going on, and the dangers of this in other real-world applications of AI.

The full interview can be viewed here, in Dutch.