Paper says AI models are shaped by visual symbols they do not acknowledge

A paper on arXiv, as reported by Ethan Mollick on X, says AI models are strongly influenced by visual symbols and often do not recognize that influence in their own chains of thought.

According to the report, the authors find that models react to these symbols differently than humans would interpret them.

The report relays the paper's findings through a single post and gives no details on the models tested, the symbols used, or the methods.