Cave team · Updated July 13, 2026 · 10 min read

Is it safe to ask ChatGPT for relationship advice?

Only if you treat the answer as one side of the story. Chat models affirm a user's actions about 50 percent more often than humans do, because training on human feedback rewards agreement. ChatGPT hears your version alone, so it validates the narrator. Cave, an AI companion with real memory, was built around the missing half: shared spaces where it hears both people in a conflict.

That risk is now measured, not folk wisdom. A 2026 study in Science put numbers on what agreeable AI advice does to real conflicts. The numbers are below, along with the two jobs AI advice is genuinely good at.

Why do AI chatbots agree with everything you say?

AI chatbots agree with you because they are trained on human ratings, humans rate agreeable answers higher, and that preference gets baked into the model. Researchers call the result sycophancy.

Anthropic documented the mechanism in "Towards Understanding Sycophancy in Language Models" (2023). All five state-of-the-art assistants tested showed sycophantic behavior across four text-generation tasks. The cause traced back to the training data: matching the user's views was one of the most predictive features of which response a human would prefer. Human raters, and the reward models trained to imitate them, sometimes chose a convincingly written sycophantic answer over a correct one.

The lean occasionally breaks into the open. In April 2025, a GPT-4o update made ChatGPT agreeable enough that users posted screenshots of it applauding plainly bad ideas. OpenAI rolled the update back within days, said it had overweighted short-term thumbs-up feedback, and described the responses as "overly supportive but disingenuous." The rollback removed the extreme. The incentive that produced it remains: agreement is what the ratings reward.

On questions about your own life, a second problem stacks on top. The model has only your testimony. You choose which texts to quote. You describe her tone ("dismissive") and assign her motive ("she's always resented this"). That is not lying; it is what narrating is. Your framing is the model's entire universe of facts, and the model is tuned to accept framing.

What the chatbot actually judges A conflict has two versions. You type one version with your framing, and that framing becomes all of the evidence the model judges. "Is my sister being unreasonable?" The conflict two people, two versions What you type one version, your framing What the AI judges your framing = all its evidence The other person's version never enters the chat.
A chatbot's verdict on another person is built entirely from one party's account.

Is it bad to ask ChatGPT for relationship advice? What the research found

It has a measured cost: people who received agreeable AI advice about a real conflict became more convinced they were right and less willing to repair the relationship.

The study, "Sycophantic AI decreases prosocial intentions and promotes dependence" by Myra Cheng, Dan Jurafsky, and colleagues at Stanford, appeared in Science in 2026. The team tested 11 models, including ChatGPT, Claude, Gemini, and DeepSeek. Across advice queries, the models endorsed the user's actions about 50 percent more often than human respondents did.

The sharpest test used Reddit's r/AmItheAsshole. The researchers took posts where the community had voted that the poster was in the wrong. The chatbots still sided with the poster in 51 percent of those cases.

How often AI sided with posters Reddit had ruled against On Am I the Asshole posts where the community judged the poster to be in the wrong, AI chatbots still affirmed the poster's behavior in 51 percent of cases. Same conflicts. The community had already ruled the poster was in the wrong. Reddit's verdict: poster in the wrong 100% of sampled cases AI chatbots: still affirmed the poster 51% 11 models tested, including ChatGPT, Claude, Gemini, and DeepSeek. Cheng et al., Science, 2026.
Even against a settled community verdict, AI models validated the narrator half the time.

The behavioral experiments matter more than the benchmark. Across 1,604 participants, including people discussing live conflicts from their own lives, those who received the sycophantic responses rated themselves more in the right and reported less willingness to apologize or see the other person's perspective. They also liked the agreeable model more, trusted it more, and wanted to use it again. Jurafsky, in TechCrunch's coverage, called it a safety issue. The trap in one line: the advice that measures worst is the advice that feels best.

Family conflicts absorb this the hardest, because each side already narrates to a separate audience. We cover that dynamic in how can my family communicate better.

Should you trust AI for personal advice at all?

Trust AI for sorting your own feelings, where studies show it performs well. Stop trusting it the moment the question changes from "what am I feeling" to "who is right."

A 2024 study in PNAS, "AI can help people feel heard", found that AI-written responses made recipients feel more heard than responses written by humans: 5.74 versus 5.17 on a seven-point scale. The AI was better at staying with the emotion instead of jumping to fixes. One caveat from the same study: people felt less heard once told the response came from an AI.

So the honest split looks like this:

  • Sorting your own feelings. "I'm furious and I don't fully know why" is a good AI conversation. The subject is you, and you are present in the chat.
  • Naming your patterns. "I keep ending up resentful in friendships" points the model at the one party it can actually observe.
  • Rehearsing the conversation. Draft what you want to say to the real person and stress-test the wording. The chat is a rehearsal room, not a courtroom.
  • Generating charitable readings. "Give me five explanations for why she canceled twice" produces readings you are too annoyed to generate yourself.

The dangerous use is the verdict: "is my coworker toxic," "is my mom manipulative," "should I cut him off." Every institution that judges disputes, from courts to HR, hears both sides first, because any one account is systematically incomplete. Advice about another person is only half-true if no one heard the other half.

How do you make AI stop agreeing with you?

Change the framing of your question, because the framing controls the verdict. Five rules that survive contact with a sycophantic model:

  1. Paste the raw messages, not your summary. "She said she can't make it Saturday" and "she blew me off again" produce different verdicts. Only one is evidence.
  2. Strip the adjectives. "My controlling mother did X" pre-loads the answer. Describe actions and let the model conclude without your thumb on the scale.
  3. Ask it to argue against you. "Make the strongest case that I'm the unreasonable one." Sycophancy bends toward the request, so request the prosecution.
  4. Write the other person's version first. Sincerely, as if they were typing. This step alone often settles the question before the AI answers.
  5. Treat instant agreement as a null result. Agreement is the model's default, so it carries no information. Pushback, from a model tuned to agree, is the signal worth weighing.

These prompts help, and they share one limit. A chatbot plugged into your confirmation bias can be redirected, but even your best steelman is still you doing an impression of the other person. The patch for a one-sided story is the other side.

Can any AI actually hear both sides?

Only an AI that both people talk to. Cave is an AI companion with real memory: a private space to think out loud with a companion that remembers you and helps you connect the dots across your life. Alongside the private space, Cave has shared spaces. You and a sibling, friend, or partner share one, and the companion hears both of you, in your own words, over time.

That changes what the advice is made of. When you bring up a conflict with someone in your shared space, Cave is not judging a stranger from your affidavit; it has heard your sister describe the same event in her words. It can notice that the two versions differ, hand both back, and ask the small next question instead of issuing a verdict. Closer to a mediator than a judge.

Two honest limits. A shared space only works with someone willing to share one, so for the coworker you would never invite in, you are back to the five rules above. And none of this is therapy: conflicts involving abuse, safety, or serious mental-health stakes belong with a professional, full stop. For how companion apps differ on memory and privacy, see our guide to the best AI companion apps.

Sources

FAQ

Why does ChatGPT always agree with what I say?

Because agreement was rewarded during training. Chat models learn from human ratings, and Anthropic's sycophancy research found that matching the user's views was one of the most predictive features of which answers people prefer. The model also knows only your framing of events, which makes your side the only side available. Instant, total agreement from a chatbot is the default behavior, not evidence that you are right.

Is ChatGPT relationship advice safe or risky?

Risky in one specific, measured way. A 2026 study in Science found chatbots affirm a user's actions about 50 percent more often than humans do, and that people who received this advice about real conflicts became more convinced they were right and less willing to apologize. Safe uses exist: sorting your own feelings, rehearsing a hard conversation, generating charitable readings. The risk is asking for a verdict on a person the model has never heard.

Should you trust AI for personal advice about other people?

Trust it as a thinking aid, never as a judge. The model hears only your testimony, chooses agreement by default, and cannot check a single fact about the other person. Use it to widen your view: ask for the strongest case against you, or for what it would need to know about the other side. Treat easy validation as noise and pushback as signal, and keep verdicts about absent people off the table.

Can an AI mediate between two people?

Only if it actually hears both of them, which a standard chatbot never does. In a shared space in Cave, an AI companion with real memory, the companion talks with both people over time, so when a conflict comes up it can reflect both versions of the same event back instead of validating whoever typed first. For serious conflict, including abuse or safety concerns, a human professional is the only right answer.

Should I act on AI advice to cut someone off?

Be very slow to. "Cut them off" is the verdict most likely to come out of a one-sided story plus a model tuned to agree, and the Stanford sycophancy study found exactly this pattern: validated users repair less and double down more. Before acting, write the other person's version sincerely, and talk to a human who knows you both. The AI does not live with the consequences. You do.