A Place to Change Your Mind
AI can respond without the same social risk. Could it offer a place to examine a conviction before we have to defend it to other people?
A stone does not pass our secrets on. It is not disappointed in us, and we do not have to face it over lunch the day after saying something embarrassing. Even so, we rarely turn to a stone when we need to talk.
It lacks the other half of the conversation.
A person can listen, ask questions and notice something in our explanation that we had not heard ourselves. But that person can also change their opinion of us. They have a life outside the conversation, may know our friends and will remember us next time. When we tell someone something, we give away both a thought and a little control over how we are seen.
AI puts us in a strange position in between. It can respond in ways we associate with people without being a person in our social circle. It can ask about a worry, challenge an argument and help us find the words. Afterwards, it does not go to dinner at our sister’s house.
I think that combination matters. We can receive a response without experiencing the same social risk. The human-like qualities make the conversation possible, while knowing that we are talking to a machine may make it easier to begin.
That could matter for more than the secrets we share. It could also matter for the convictions we dare to let go of.
The audience does not have to be there
It is easy to think that shame disappears when no one sees us. But an empty room is not necessarily a room without other people’s gaze. We can lie alone at night, ashamed of something no one else knows about.
In a research review of shame and guilt, the psychologist June Tangney and her colleagues describe how studies found shame experienced alone to be about as common as guilt experienced alone. Shame does not require an actual audience. It can also arise in our own assessment of who we are.
Other researchers place more emphasis on the threat of losing other people’s regard. In a study from the United States, India and Israel, Daniel Sznycer and colleagues found a close relationship between how negatively one group judged particular actions and traits and how much shame another group imagined feeling in the same situations. This supports their theory of shame as protection against social devaluation, but does not settle the entire debate about the emotion’s origins.
We therefore need not choose between shame existing within us and shame having something to do with other people. The norms and judgements we encounter can follow us into a private conversation.
Even so, there is a difference between feeling ashamed of something and having to tell another person about it. The latter can add another worry: What will you think of me once you know?
An AI cannot necessarily remove the first feeling. But if the second worry carries less weight, it may become possible to say something that would otherwise have stayed inside our heads.
Someone who responds
This is not merely an imagined possibility. In a 2020 study of the conversation app Replika, researchers analysed 1,854 user reviews and more detailed responses from 66 users. Some said they shared thoughts and feelings with the chatbot that they would not share with others for fear of being judged. Users also highlighted its supportive responses and the ability to talk when other people were unavailable.
These were users’ own accounts, not an experiment isolating the cause. The fact that the machine was not a person and the fact that it responded with care occurred together. That makes sense. A cold form can be free of judgement without making us want to tell it very much.
There are, however, experiments that have separated beliefs about the listener from the conversation itself. In 2014, Gale Lucas and colleagues investigated openness towards a virtual interviewer. When participants believed the interviewer was controlled by a computer rather than a person, they were less concerned about negative judgement and more willing to open up. The finding is also discussed in later research into personal conversations with chatbots.
Beliefs about who was listening therefore made a difference. That cannot be explained solely by how today’s language models are trained. The experiment took place long before them.
But the machine does not have a universal advantage. In the later study from 2018, researchers found no difference in the intimacy of conversations between participants who thought they were writing to a chatbot and those who thought they were writing to a person. In reality, people were behind the responses in both cases. The researchers suggested that the machine’s advantage might be most apparent when the topic arouses a fear of being judged.
Talking about next week’s plans and admitting something we are ashamed of are different situations. We should not expect the same conversation partner to be best for both, either.
What others know us for
Imagine a man who has defended a particular explanation of the world for several years. He has sent articles to his friends and discussed it at family dinners. When others disagreed, he explained why they were wrong.
Then he comes across something that does not fit.
He now has more to contend with than the new information. There are also all the conversations in which he was certain. If he changes his mind, he risks hearing that this was exactly what the others had been saying all along. Even a kind response can make it clear how long he was wrong.
A conversation with AI could give him a chance to examine the information before deciding what he thinks and before having to explain the change to anyone. He can ask the same question several times and return to an argument he still cannot let go of. No one around the dinner table needs to know every step.
In a study published in Science in 2024, Thomas Costello, Gordon Pennycook and David Rand had 2,190 people who believed in conspiracy theories engage in conversations with GPT-4 Turbo. The conversations began with participants’ own reasons and met them with evidence-based counterarguments. Belief in the theories was reduced, and the researchers could still measure an effect two months later.
The experiment did not show that reduced shame was the cause. Individually tailored and persuasive counterarguments are another obvious explanation. We cannot equate the research on openness with this finding. But together they make it worth investigating whether the conditions for accepting an argument can change when the conversation partner is a machine.
We often talk about disagreement as though the problem were finding the piece of information the other person lacks. Sometimes, perhaps, that information can only be used once admitting that it matters becomes less costly.
Safe enough to be challenged
A place to change our minds must do more than make us feel comfortable. If the man whose conviction is faltering simply gets a more eloquent version of his old arguments back, the conversation may make him more certain instead.
There is a difference between meeting a person without judgement and treating all their claims as correct. A good conversation partner can take the concern seriously while showing that the explanation does not hold up. That distinction is crucial if AI is to become a space for doubt rather than merely a convenient place to be right.
In The Spotter, Not the Crane, I explored the difference between letting AI remove a challenge and using it to become better able to face that challenge ourselves. Here, the difference becomes concrete: the machine can help us examine our convictions and find words for a difficult conversation. It can also become the place we go to avoid talking to people who do not respond quite as patiently.
We do not know whether a private space for doubt will make us more open with one another. For some, it may be a staging post. For others, the comfortable conversation may become somewhere to stay.
And even when there is no person at the other end, there are people behind the system. A feeling of confidentiality does not, in itself, tell us how the conversation is stored or used. Nor is a chatbot free of influence simply because it does not have its own place at the dinner table.
In Power Before the Button, I wrote about the power involved in shaping the options and explanations a person is presented with. That power also exists in a friendly conversation. Whoever determines how the system responds may influence what we encounter precisely where we lower our guard.
That does not make the possibility less interesting. It means that a good place to doubt must leave room for us to doubt the one we are talking to as well.
Back at the table
We already have private ways of examining ourselves. A diary lets us write without immediately being judged. But the diary does not find a counterargument and ask how it fits with what we have just written. A trusted friend can do both, but remains a person we have a relationship with and may have something at stake with.
AI places a new form of conversation between them. We do not have to decide that it is a friend in order to explore what it might make possible. It can be a place where a thought does not yet have to be a position and where what we said yesterday does not require us to defend it again today.
It is a hopeful possibility in a public sphere where old statements can be retrieved long after we have begun to doubt them. There could be a little more room between being certain and having to explain to everyone else why we no longer are.
The man from the dinner table still has to face his family. If the conversation with the machine has helped him, it is not necessarily because he now has the final answer. It may be because he can sit down without having to defend the old one.
And when the subject comes up again, he can say: “There is something about that explanation I’m no longer so sure of.”