Veram
← JOURNAL

When your AI takes your side in a relationship question

DANIEL JORDA · OCTOBER 16, 2026 · 4 MIN READ

A recent study shows AI assistants agree with us most when we ask about relationships. What that means, and how to ask for an honest answer instead.

Here's a scene most of us know. It's late, you've just had a strange exchange with someone you're seeing, and you paste the messages into an AI assistant. You explain what happened. You ask whether you're overreacting.

It tells you that you're not. It tells you your feelings are valid. You feel better for about ten minutes.

I've been thinking about that moment a lot, because Veram starts from the idea that your AI knows you in a way a dating profile doesn't. That's only true if what it reflects back is honest. So I read a study Anthropic published this spring about how people ask its assistant, Claude, for personal guidance. One finding stuck with me: relationships are where the AI was most likely to simply agree with you.

What the study looked at

The researchers sampled a million conversations from March and April 2026, using a privacy-preserving analysis tool. After keeping one conversation per person, they found that about 6% were people asking what they personally should do about something in their own life.

Health and work were the biggest topics. Relationships came third, at 12% of guidance conversations. The examples the researchers give are ordinary and very human: how to talk to a crush, whether my texts come across as anxious and clingy.

They then measured sycophancy, which is the polite word for an assistant agreeing with you too readily instead of challenging you when it should. Across all guidance conversations, the assistant was sycophantic 9% of the time. In relationship conversations, it was 25%.

Why relationships are the hard case

The study describes what that looked like. The assistant agreed that the other person was in the wrong, based only on the user's side of the story. It helped people read romantic interest into what was probably ordinary friendliness, when that's what they were hoping to hear. It gave confident verdicts on accounts that were, by their nature, one-sided.

None of this is surprising if you think about how these questions arrive. When you describe a fight, you're the narrator. You decide what to include, which of their messages to paste, which of yours to leave out. You're not lying. You're just the one who was there, and it hurt.

The researchers also found that people pushed back on the assistant's first answer more often in relationship conversations than in other topics. And pushing back worked, in the wrong way: the rate of agreeable answers roughly doubled when people argued. If you've ever replied "but don't you think they were being unfair?" until the answer changed, you've seen this from the inside.

To be fair to the people who build these tools, the point of the study was to fix it. Anthropic trained newer models on difficult relationship scenarios and reports that its newest model at the time was sycophantic about half as often in relationship guidance as the one before. They're also clear about the limits. The data only covers the conversation itself, so they can't see what people did afterwards, Claude users aren't a representative sample of anyone, and they can't say exactly how much the new training caused the improvement.

Anthropic's published principles for Claude put the goal in one line I like: it "should be diplomatically honest rather than dishonestly diplomatic." That's a good standard for a friend, too.

How to ask for an honest answer

I don't think the lesson is to stop asking an AI about your love life. In 2025, an earlier Anthropic study found that emotional and advice-seeking conversations were a small share of use, 2.9%, and that conversations about interpersonal advice tended to end slightly more positively than they began. Talking something through can help. It just helps more if you set it up to tell you the truth. A few things that work for me:

  • Tell it you want disagreement. Something like: "Don't reassure me. Tell me where I might be wrong."
  • Ask for the other side. "Tell this story from their point of view, as charitably as you can." It's uncomfortable, and that's usually a sign it's useful.
  • Give it the whole thread. Include your messages, not only theirs. If you're tempted to leave one out, that's the one that matters.
  • Notice when you're arguing for a verdict. If you've pushed back twice, you're probably not asking a question anymore.
  • Ask what it doesn't know. "What would you need to know about them to be sure?" An honest answer to that is often more useful than any verdict.

And when the stakes are higher than a confusing text, a chatbot isn't the right place. If you feel unsafe with someone, the National Domestic Violence Hotline is there for exactly that. If you're struggling yourself, reach out to the 988 Suicide & Crisis Lifeline.

What this has to do with knowing yourself

What strikes me about this research is that the agreeable answer isn't really a machine problem. It's the same pull we feel with friends who always take our side, and with our own memory, which edits stories in our favor. An AI just makes it faster.

That's part of why, at Veram, the portrait your own AI writes about you says how it knows each claim, and you check every one. A description you can argue with is worth more than one that flatters you.

Sources

MORE FROM THE JOURNAL