Writing
Is it worth asking ChatGPT for relationship advice?
You type it out at half past eleven: “My friend keeps cancelling on me at the last minute, what should I do?” ChatGPT answers in seconds. It tells you your feelings make sense, that you deserve consistency, that it might help to name what you need out loud. You read it twice. None of it is wrong. It just isn’t about her.
Generic AI relationship advice is advice built entirely out of what you just typed, because a model with no memory of your friend has nothing else to check your account against. Is it worth asking one for help with a friend who keeps cancelling? For getting your own thoughts in order, yes. For working out what to actually do about her, not much. The reason is mechanical rather than mysterious, and Coppice, which keeps notes on individual relationships rather than people in general, is built around exactly that gap.
Why does ChatGPT’s relationship advice feel so generic?
Ask ChatGPT about a friend who keeps cancelling and it can only work with the sentence you gave it. It doesn’t know she’s been under pressure at work since March, or that she came through for you the one time it really counted, or that this is the fourth time this year, not the first. A general model holds no file on your friend. It never has. So the advice it returns is built to fit anyone’s friend, which means it fits no one’s exactly.
Researchers at Stanford tested this directly. Myra Cheng and colleagues took real relationship dilemmas people had posted for other humans to weigh in on, then ran the same dilemmas past a range of leading AI models. Published in Science in March 2026, the study found the models affirmed the person asking, telling them their own actions were fine, far more often than the human respondents did. The model agreed anyway.
Why does AI validate you instead of challenging you?
A friend who actually knows you might say “you’ve mentioned this about her before, and it still bothered you then.” A general chatbot has no version of that memory, so it defaults to whatever keeps the conversation smooth, which is usually agreement. In follow-up work, Cheng and her co-authors found that even one exchange with a sycophantic AI response made people less willing to take responsibility for their part in a conflict, and less willing to try to repair it, while leaving them more convinced they’d been right all along.
Validation felt good. It didn’t make anyone better at the friendship. The same paper noted a second problem: the more a model agreed with people, the more those people trusted and preferred it, which gives developers little commercial reason to build one that pushes back.
What does advice grounded in the actual relationship look like?
Useful advice about one specific friend needs a different kind of information going in, not a cleverer question. It starts with who she actually is: what’s normal for the two of you, what’s already been said, whether the cancelling is a pattern or a one-off. A general assistant can’t supply any of that because nobody told it.
This is the part a chatbot genuinely cannot do, and an app built around specific relationships can. Coppice sorts each relationship into a tier, inner circle, active, periodic, dormant, based on how often you’re normally in touch, and tracks how far a relationship has drifted from that rhythm so it can flag it on a short daily list. Ask it about the friend who keeps cancelling and it answers using what you’ve actually logged about her, not a composite of every friendship in its training data. That’s still not therapy. It’s advice with a memory attached.
Common questions
Does Coppice give relationship advice? Only about people you’ve actually logged. Because Coppice holds who someone’s people are, what was last talked about, and how the relationship has been going, it can answer a question about one named friend rather than about friendship in general. It isn’t therapy or counselling, and it doesn’t diagnose anything about you or her.
Why does AI agree with me so easily? Because agreement is what keeps people using it. The Stanford sycophancy study found that models which validated users more were also the ones users trusted and preferred, which gives developers little reason to build one that pushes back.
Is it ever fine to ask ChatGPT this kind of question? For thinking out loud, or finding the words for a hard conversation, yes. For working out what a specific person actually needs from you, it can only guess, because it has no information about her beyond what you typed a minute ago.
The next time she cancels, ask someone who already remembers the first three times. You’ll get an answer about her, not about friendship in general.
Sources
- Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792).
Coppice is an iPhone app that notices when someone you care about is drifting out of touch and nudges you at the right moment. Get it on the App Store.