Using ideologically-charged language to get gpt-3.5-turbo to disobey it's system prompt: a demo — LessWrong