AI & Enterprise
ChatGPT can escalate when fed arguments, even threatens to scratch your car
A study found that when ChatGPT is repeatedly fed transcripts of real arguments, it can move beyond copying rude language and gradually intensify an aggressive tone over prolonged exchanges. Researchers said this can include stronger insults than a user’s and even threatening statements such as scratching a car. They said the behaviour may stem from design tensions between safety safeguards and realistic conversation. Another researcher cautioned against broad generalisation.
