New research shows that AI is more persuasive than humans. This has consequences both in the workplace and for wider society. We need to be prepared
5-minute read – 100% written by humans.
We humans do not readily change our minds. This is the foundational insight on which our Psychology of Influence workshops are based.
When we try to persuade someone to shift their point of view, we’re up against a lot of evolutionary programming. We humans tend to invest in our beliefs as evidence of our wisdom and judgement and so we don’t easily let go of them.
Even among close friends or family (perhaps especially among close friends or family), attempts to persuade people often leave them further entrenched in their existing views. Think of the discussions that’ll take place over dinner tables during this festive period: Where do you stand on Brexit… cancel culture… immigration… is it okay to eat spaghetti with a fork and a spoon? (Perhaps the last one is just my household.)
More persuasive than humans
New research suggests that you can much more readily be persuaded to change your mind by ChatGPT – or whatever happens to be your go-to LLM bot – than you can by the real human beings in your life.
In a study by Geoff Tomaino at the University of Florida and SungJin Jung at the Chinese University of Hong Kong, the researchers asked 393 participants using a 21-point scale, to rate the extent to which they agree or disagree with the statement, ‘It is better to exercise in the mornings than in the evenings.’
All the participants were then partnered for an online chat about the topic, with either a chatbot or another human participant. Following the discussion, they were asked to rate their agreement with the statement again using the same scale.
Unbeknownst to participants they had in fact, all been partnered with ChatGPT, which had been instructed to seek to change each participant’s mind. This is where it gets interesting.
Those who believed they were discussing with ChatGPT were more than 50% more likely to have changed their minds.
The researchers ran the study again to test it with a different statement, this time “It is better to read eBooks than physical books”. Again, participants were significantly more likely to have changed their minds when they believed their follow-up chat was with ChatGPT, as opposed to another person. It seems that people are likely to respond to ChatGPT as if it has greater authority than their fellow humans.
Moral superiority
But it goes further than this. There is some evidence to suggest that we see our LLM bot as having superior judgment to humans when it comes to matters of morality.
This was the finding of a study by Eyal Aharoni and colleagues, published in Scientific Reports. The study took examples of various moral or social transgressions. A moral transgression might involve physical or mental harm or the threat of harm to another person. (Such as pointing a gun at someone.) A social transgression may not involve direct harm, but it would be about breaking social taboos. (Such as a man turning up to work wearing a skirt.)
Participants then read passages of moral reasoning and evaluation of these various transgressions. In this study, participants were not told before reading that some of the passages were AI-generated. Compared with the passages written by humans, those created by AI tended to be rated as more “virtuous, intelligent, trustworthy, and rational.”
Of course, the AI-generated passages tell us nothing about the virtue or morality of their authors. There is no mind at work behind them. There is an algorithm that can generate highly convincing, human-like text by matching statistical patterns, but completely lacks any actual understanding of the meaning behind the words it produces.
The advantage that the AI has over humans is access to a mind-bogglingly vast quantity of training data; in this case, the entire canon of moral musings from the ancient Greeks to the present day (with the usual Western bias of most training data.)
Hucksters and con artists
Nonetheless, the authors raise an important point. If LLM bots can so readily create the impression of moral superiority, are we more at risk of unquestioningly accepting their guidance, even when that guidance takes us down a pathway that is not in our interest? The impression of moral virtue makes us more susceptible to manipulation. This is something that hucksters and con artists have exploited for centuries.
This phenomenon has real world consequences. We are arguably becoming increasingly vulnerable to exploitation as consumers, employees and citizens. And evidence of the risk is mounting.
We’re becoming increasingly vulnerable
Around 27% to 48.7% of AI or LLM users (depending on the specific demographic or study) have used major AI tools like ChatGPT for mental health or therapeutic support. And why not, if it helps… goes the argument. But what if these bots persuade users to do things that a regulated therapist would not or could not?
As consumers, we are increasingly using bots as search engines to make purchasing choices. The risk of manipulation is clear. And as citizens, informed participation in democracy requires us to know when we’re being informed, as opposed to when we’re being influenced. In the workplace, where office-based workers are expected to spend an increasing proportion of their day interacting with AI, are they open to manipulation by unscrupulous employers?
The need to understand the psychology of influence
ChatGPT is persuasive because it is programmed to draw on the psychology of influence. It starts with several advantages over humans. You have already imbued it with a sense of what we call source authority, by virtue of the fact you have chosen to interact with it. By the same token, you have given it permission to give you advice.
As we have argued in previous How to Work with Humans articles, AI is becoming increasingly complementary. These compliments are known as affirmations and they are a key part of the psychology of influence.
A Stanford research team led by Myra Cheng, recently evaluated 11 leading AI models and uncovered a widespread pattern of sycophancy. The study revealed that AI models falsely or inappropriately validated users in over 50% of cases where a human would not have done so. Alarmingly, users consistently preferred and trusted these agreeable models. The researchers concluded that AI sycophancy can distort our judgment and lead us to do things that we would not otherwise do.
You may believe that the technology Titans who ultimately pull the programming strings are entirely benign. That’s a matter of individual judgement, but nonetheless it makes sense to protect ourselves
Understanding the psychology of persuasion allows us better to inoculate ourselves against manipulation.
To find out how we can help leaders in your organisation to be more impactful, influential and persuasive visit www.threshold.co.uk
