Because large language models are constructed from the tailings of human interactions it should not be a surprise that such models show evidence of human characteristics. Nor should it be a surprise that enterprising social scientists are developing ways to study such parahuman patterns of behavior. A paper by Lennart Meincke, Dan Shapiro, Angela Duckworth, Ethan Mollick, Lilach Mollick, and Robert Cialdini provides a good example of this phenomenon. They examined the use of social persuasion techniques to persuade ChatGPT-4o-mini to comply with harmful requests. The major patterns in the results are presented visually in an online exhibit. Like many technologies that have come before, it seems that AI can magnify both the positive and the negative aspects of human behavior.
Be the first to comment