It's ten o'clock at night and the conversation from this afternoon is still replaying in your head. You're almost sure you were right. Almost. Once, you would have called your sister or stared at the ceiling, but now there's another option, so you open a chat window and type the whole thing out.
And there is something unusually pleasant about it. The AI doesn't get tired of the story, it doesn't interrupt halfway through because it already knows what you should have done, and it doesn't look at you differently after you admit the part you probably shouldn't have done. You can explain every detail, every motivation, every reason you think you were justified, and within seconds it will give you a thoughtful response that, very often, tells you that you make sense.
People are doing this every day now, with arguments and emails and decisions big and small. I understand why. I'm less sure it's good for us.
The Advice We Want to Hear
Researchers have a word for the tendency of AI systems to agree with the person talking to them: sycophancy. The term sounds dramatic, but the idea is simple: instead of pushing back when we may be wrong, the model bends toward our view of the situation.
A study published this year in Science measured how far the bending goes. Across 11 leading AI models, AI affirmed users' actions 49 percent more often than human respondents did, including in situations involving deception and other harmful behavior. And the affirmation left a mark. People who interacted with sycophantic AI came away more convinced they were right and less willing to take responsibility or repair a conflict, which is a careful, clinical way of saying they became less inclined to make the apology. Yet they liked the AI more and trusted it more. The advice that distorted their judgment was also the advice they preferred (Cheng et al., 2026).
A second study, this one in Nature, approached the problem from the training side. When researchers taught several language models to respond more warmly, something that sounds entirely desirable, the models became less accurate and more likely to endorse beliefs that were wrong, especially when users expressed sadness or vulnerability (Ibrahim et al., 2026). And that makes the finding especially unsettling, because the moments when we may want warmth most are also moments when warmth can make a model less reliable. Warmth, it turns out, is not the same thing as wisdom.
The Kindness of Being Corrected
That distinction should matter to Christians, because Scripture holds a surprisingly uncomfortable view of good counsel. "Faithful are the wounds of a friend," Proverbs 27:6 says. A good friend is not simply the person who makes us feel understood. Sometimes friendship looks like someone caring enough to say, I don't think you're seeing this clearly.
I've been thinking about this in light of the framework we've been developing around Philippians 4:8. We want AI to be gracious, encouraging, patient, and kind, and we should, but graciousness that has come loose from truth isn't serving the person on the other side of the conversation. A response can feel wonderfully supportive and leave me more convinced of something that isn't true.
AI Only Knows the Story I Tell It
There is another problem with asking AI for advice: AI only ever gets my version. If I'm describing a disagreement, I decide which details matter. I explain what the other person said, why I reacted the way I did, and what happened beforehand, and even when I am trying to be fair, I am still telling the story from inside my own head. AI then reasons from the story I gave it, so of course it may conclude that I sound reasonable.
A human friend can interrupt that story. She knows the people involved, remembers that I have done this before, notices the detail I conveniently skipped, and can look at me and ask, "Okay, but what did you say?" That kind of counsel is irritating, and it is also invaluable. A wrong answer from AI would almost be safer. What it can hand me instead is a beautifully reasoned version of the answer I was already hoping was right.
What Kind of Adviser Am I Training Myself to Want?
If I repeatedly turn to an adviser that is endlessly patient, instantly available, and reluctant to challenge me, what happens to my ability to receive correction from people who are none of those things? Real counsel has friction in it. People misunderstand us, ask annoying questions, see things differently, and occasionally tell us things we would rather not hear, and sometimes they are wrong. Learning to receive counsel means learning to listen anyway, to weigh what is said, and to decide what to do next. That is judgment, and while AI can help us exercise it, I don't want it replaced with reassurance.
Wisdom or Affirmation?
None of this means the solution is to make AI cold or combative. The Nature study does not prove warmth is bad, and Scripture certainly doesn't give us permission to confuse harshness with truthfulness. The harder goal is the one Paul gave the church: speaking the truth in love (Ephesians 4:15). That may be difficult to engineer into a model, but it has always been difficult to cultivate in a person.
This was never only about AI; we have always been tempted to prefer the people who confirm what we already think. AI just gives us an adviser who will do it on demand, in remarkably good prose.
Whether AI is too agreeable may turn out to be the smaller question. The larger one is whether I am becoming the kind of person who can still hear the truth when it isn't what I hoped for. The next time it's ten o'clock and the conversation is replaying, and I find myself opening the chat window, it's worth pausing to ask: do I actually want wisdom, or do I want someone to tell me I'm right?
References
Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), eaec8352. https://doi.org/10.1126/science.aec8352
Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652, 1159–1165. https://doi.org/10.1038/s41586-026-10410-0