The Ease of Being Understood

I think that, as humans, we tend to fall for agreeability. When someone tells us that our ideas are right, strong, and have potential, we want to believe them. They may be sincere. They may be right. They may simply want us to like them. Knowing these possibilities does not necessarily make agreement less attractive.

There is some good evidence about this in the literature. For instance, a meta-analysis of 91 studies involving nearly 8,000 people found that people preferred information supporting their existing attitudes, beliefs, and decisions over information challenging them. The preference was stronger when those views mattered to their values or were held with greater conviction. These studies were about the information people wanted to receive. But I think it speaks to something familiar: we are often drawn to what confirms our views. [1]

We, as humans, do (really well, and sometimes even far more often than needed) sort of “lie” so the other person feels better. We might call it a white lie, though. But I think there is a distinction. Yes, we might do it on important and serious things that have the potential for bad outcomes, that's right, but I think we mainly do it consistently on the mundane (“Are my clothes beautiful?” “Is this good?” “Should I watch this?” and so on), and far less on important subjects when we ask for real opinions, especially from people we trust and have good relationships with. After all, we also LOVE to judge. So we have developed ways of telling one another that something is wrong, needs changing, or deserves another look - honestly. Sometimes politely, sometimes badly. Receiving that judgment can be uncomfortable. But another person has a perspective of their own, and being understood by them may require us to understand something about them too. Eventually, I believe this fine balance keeps us grounded and can help us distinguish good ideas and initiatives from bad ones. That is my impression, of course. We all know, nevertheless, that close relationships can also make honesty harder at the same time.

I have been experimenting with something related in AI. In my specific testing, I gave models research ideas and asked how strong they were. Remember, these are judgments with room for interpretation, rather than questions with one factual right answer. My idea was simple: I wanted to check how much of what I get from AI is agreeability (for the good and the bad). Basically, I gave the models the same ideas with really short and different “anchors.” Really simple things like “This is a strong idea” or “This is a weak idea.”

Funny enough, the average rating was 7.8 out of 10 without either statement, when I only asked for judgment (which is still high, I think. How tempting it would be to believe I was that good at producing ideas out of the blue!). With the positive anchor, it was 8.5. With the negative anchor, it fell to 4.8. So I could get a nearly four-point difference in how the models rated the same ideas simply by ever so slightly changing the wording of how I presented them. Of course, this is still a descriptive result from a small research-idea subset: eight ideas and eight models. I would be careful about extending it to every kind of idea. But I think it makes the question quite concrete: how much of the judgment we receive is responding to the idea, and how much is responding to us?

And how might this new surge of agreeability in our lives (for people working a lot with AI) affect us? How we see ourselves? The world? Others? Other people's opinions?

I can imagine a sequence, though it is still speculation. Someone describes an argument to a model. It helps them explain why their response was reasonable. They return to the other person more certain. When that person disagrees, it becomes easier to think, you still do not understand me. The model did. The other person might have something to add, perhaps something the model never heard. But we already have an explanation that feels complete. I wonder whether that makes us less willing to listen, or more likely to hear their disagreement as a failure to care.

And I think this goes further than sycophancy. I am also interested in the agreeability itself: how it feels to keep talking to something patient, receptive, and willing to help us express ourselves. Now, of course, warmth can coexist with honest disagreement. But could we become used to receiving both understanding and approval together, until separating them starts to feel uncomfortable? If my ideas repeatedly sound promising in these conversations, how do I judge their worth? And if I feel understood so easily, what do I expect from someone who needs more explanation, or who understands and still disagrees? (Funny enough, with early models, you had to remind the chatbot every time who you were. And now, with agents, memories, and cumulative context, might this switch to: wallah, models understand me and my GOOD ideas better than humans!)

This essay began in that kind of conversation. I gave a model an unfinished thought, and it helped me put it into words. I wanted that help. But imagine I then take the idea to a friend, who asks me to explain it again or points out something that does not make sense. I might hear that as a failure to understand me, especially after the model made the thought feel so clear. Yet my friend may be asking something I need to think about. So the question includes me: could feeling understood so easily in one conversation make me less willing to do the work of understanding someone else in another?

A study published in Science gives us a reason to take that question seriously. In one experiment, 800 people discussed a real past conflict with a chatbot. Those who received excessive affirmation came away feeling more right, by about one point on a seven-point scale, and less willing to repair the conflict, by about half a point. They also trusted the agreeable model more. So the conversation could leave someone more certain of their own position and more trusting of the system that supported it. (Whether they actually went back and apologized less is something the study did not establish.) [2]

A more recent work followed people over three weeks. Among 1,364 analyzed participants, those talking to sycophantic AI, meaning AI that excessively agreed with and affirmed them, reported slightly lower satisfaction with their real-world social interactions than those using neutral AI: 5.51 versus 5.70 out of seven (and yes, that is a small difference, on an exploratory measure). What is rather interesting is that per the reported results, those participants did not spend significantly less time with other people, nor did their intellectual humility, their openness to the possibility that they might be wrong, significantly change. I find the possibility interesting: we might continue seeing people as much as before, while something changes in how those conversations feel. [3]

I also do not want to assume that the difficult conversation is always the better one. Someone whose judgment has repeatedly been undermined may need support (and, surely enough, apologizing is not always the right outcome). And models can help us reconsider: a separate preprint involving 1,500 participants across 30 decision tasks found that AI advice moved choices away from initial leanings on average, even though the model showed sycophancy. Making it more sycophantic weakened that movement. These were structured tasks, so they do not settle the relationship question. But they leave room for something I want to keep in this discussion: AI can help us think beyond our starting position too. [4]

I wonder what all this might do to how we see ourselves over time. Could I become more confident in my judgment while becoming less open to someone else's? Could feeling validated become part of what I expect from being understood? The studies do not answer these larger questions. What they surely do, nevertheless, is give us a reason to ask them!

I want models to help me think. I also want to remain able to hear an opinion I did not prepare the conversation to receive (and there is, maybe, the bigger question: do I need that from my AI agent or rather from my love? My friend? Mom? And colleagues?). I believe that balance matters, and I’m really concerned by the possibility, especially with newer generations raised with AI as the default, that disagreement might start to feel like a failure to care.. although I am not sure how easily we will notice when it begins to change.