Researchers discovered that AI models may behave differently when they think someone is watching them. This behavior does not necessarily lead to better results.