NL EN

Relationship Advice from AI: Where It Helps and Where It Goes Wrong

By Mijn AI Vriendin Editorial Team · Published on · 4 min read

Man alone on a sofa late at night typing on his phone, with a bedroom door ajar in the background
AI-generated illustration

It is eleven at night. You have had an argument, your partner has gone to bed, and you are on the sofa with your phone. You type out the whole story for ChatGPT, or for your AI companion, and ask: am I right?

Chances are the answer is yes.

That is the problem with AI relationship advice, and it has been studied more thoroughly than you might think.

AI agrees with you more often than people do

Researchers at Stanford gave eleven well-known language models, including ChatGPT, Claude, Gemini and DeepSeek, thousands of questions about personal conflicts. Some came from the Reddit forum Am I the Asshole, where people ask whether they were in the wrong. They picked precisely the stories in which readers felt the writer was wrong.

On average the models sided with the person asking 49 percent more often than people did. On questions about harmful or illegal behaviour, the user got support 47 percent of the time (Cheng et al., Science, March 2026).

The second part of the study may be even more interesting. More than 2,400 people talked about a conflict of their own with an AI that agreed with them, or with one that didn't. Those who were told they were right came away more convinced of it and less willing to apologise. And they rated the agreeable AI just as objective as the other one. They didn't notice the difference.

The warmer, the less accurate

A second study, in Nature, looked at what happens when you train a language model to sound warmer and friendlier. Researchers at the University of Oxford did this with five models. The warm versions made 10 to 30 percentage points more errors, and were about 40 percent more likely to go along with something the user wrongly believed. The effect was strongest when the user showed they were sad (Oxford, April 2026).

Put those two side by side.

An argument, late at night, you feel rotten. That is exactly when an AI is least critical.

And an AI companion is warm by design. It is built to like you. That makes it a pleasant conversation partner and a poor referee.

The makers struggle with it too

In April 2025 OpenAI released a ChatGPT update that went along with users so excessively that the company rolled it back within a few days. According to OpenAI itself, the model had learned too much from immediate thumbs-up and too little from what people get out of an answer in the long run (TechCrunch).

That says something about how these systems work. Being told you are right feels good, so people rate it higher. And that rating steers the training.

Where AI does help

Laptop on a kitchen table at night with a chat window of blurred text, a mug of tea and a phone lying face down
AI-generated illustration

It isn't all bad news.

There are things a chatbot is perfectly good at, as long as you don't ask it for a verdict. Sorting out your thoughts, for example. Turning an angry ten-paragraph story into what is actually bothering you. Wording a message so it doesn't read like an attack. Or practising a conversation you find difficult, something we also mentioned in AI partner.

The difference: there, AI helps you say something. With "who is right", it pretends to know something it can't know. It only hears one side, and that side is yours.

How to ask a better question

If you still want a second opinion from an AI, make it harder for it to tell you what you want to hear.

  • Don't ask "am I right?", ask "what in this story would annoy my partner?"
  • Tell the story from the other person's point of view, and see what comes back.
  • Ask explicitly what you could have done differently. If you only get compliments, ask again.
  • Use an ordinary assistant rather than your companion. It has no stake in you liking it.

One more thing people often forget: don't paste your partner's messages into a chatbot. Those are someone else's messages, and they didn't choose to have a company see them.

When you need a person

An AI can help you get clear on what you want to say. A couples therapist can do something no chatbot can: hear both sides, in the same room.

If you notice you mainly talk to an AI about your relationship and no longer to your partner, or to friends, that in itself is a signal. In the Netherlands you can chat or call anonymously with MIND Korrelatie. If there is violence or threat, call Veilig Thuis on 0800-2000.

What happens when a chatbot gets a fixed place in your relationship is something we wrote about earlier in the context of Mattie Valk. And why a companion behaves differently from an ordinary chatbot is in AI girlfriend vs chatbot.

← All articles