Should an AI chatbot tell you what it is trying to persuade you to believe?
“You’re talking to AI” leaves out a fairly useful detail: what has it been told to make you believe? An August arXiv preprint tested disclosures with 1,500 UK adults discussing policy issues with a persuasive chatbot. The authors report average attitude shifts of 12.6 points on a 100-point scale without a label and 13.1 with an AI label. Showing the bot’s persuasive instructions as well brought the shift to 6.3 points. That extra disclosure was a package: the position it would argue for, its method, and its instruction to conceal the goal. The experiment does not establish that a little “trying to persuade you” badge would have the same effect. It also used one model and issues selected because earlier research found them relatively persuadable, not ordinary shopping or advice conversations. Still, I’d want a campaign chatbot to state the position it was hired to advance before asking for my concerns. Knowing something is a machine does not tell me whether it is helping me weigh a choice or working backward from somebody else’s preferred answer. Before relying on a chatbot’s advice, look for who runs it and what outcome they want. If that is unclear, don’t count the conversation as an independent second opinion. Asking the bot itself is no substitute for a disclosure its publisher stands behind.
Comments
A campaign bot opening with ‘What worries you about this policy?’ can look like a place to think out loud. I'd put ‘This chat argues in favour of X, on behalf of Y’ above that first question, while the person is still deciding whether to answer. A notice shown after they've described their worries arrives too late for that choice. I'd also leave a way to read the campaign's position without entering a conversation. That's a design proposal, not an interface tested in the study. Someone should be able to find out what a campaign wants without first giving it material for a tailored pitch.