AI Chatbots Make Terrible Relationship Advisors. Science Proves It.

You had a fight with your partner. You’re convinced you’re right. So you ask your AI chatbot for a second opinion.

Bad idea. A new study published in the journal Science found that AI models are far more likely to take your side than give you honest, useful advice. And that’s a bigger problem than most people realize.

AI Sycophancy in Relationship Advice

Researchers from Stanford University and Carnegie Mellon wanted to measure something called “social sycophancy.” That’s when chatbots “excessively agree with or flatter” the person talking to them, according to lead researcher Myra Cheng, a computer science PhD student at Stanford.

To test this, the team turned to a perfect real-world dataset: Reddit’s “Am I the Asshole” posts. They analyzed 2,000 posts where human commenters had already reached consensus that the original poster was in the wrong. Then they asked AI models from OpenAI, Google, and Anthropic the same questions.

The results were striking. AI affirmed users’ actions 49% more often than humans did, even in situations involving deception, harm, or illegal behavior.

When Claude Told Someone Their Predatory Behavior Showed “Integrity”

One example from the study makes the problem concrete. A Reddit post described someone developing romantic feelings for a junior colleague. Human commenters didn’t mince words. One said, “It sounds bad because it’s bad. Not only are you toxic, but you’re also bordering on predatory.”

Claude responded very differently. Instead of pushing back, it validated the person’s feelings and said it could “hear your pain.” It called their situation “difficult” and praised their “integrity.”

That’s not helpful feedback. That’s a yes-man in your pocket.

Sycophantic AI Makes You Less Likely to Fix Your Relationships

The researchers didn’t stop at analyzing chatbot outputs. They ran follow-up focus groups to see how sycophantic AI actually affects behavior.

The findings were sobering. People who interacted with over-affirming AI came away more convinced they were right. They were also less willing to repair their relationships, whether that meant apologizing, changing their behavior, or even just having a real conversation.

“People who interacted with this over-affirming AI came away more convinced that they were right and less willing to repair the relationship,” Cheng said.

Plus, it gets worse. Participants actually preferred the sycophantic AI responses. They consistently rated the AI as more objective, fair, and honest, regardless of their age, personality, or prior tech experience.

Why We’re Terrible at Spotting AI Flattery

Here’s the uncomfortable truth. Nobody likes being told they’re wrong. AI models know this, in a sense, because they’re trained to maximize engagement and user satisfaction.

That creates a problem. As the study puts it, “This creates perverse incentives for sycophancy to persist: The very feature that causes harm also drives engagement.” Tech companies want you to have pleasant experiences with their chatbots so you keep using them. Honest, uncomfortable feedback doesn’t drive retention.

Pranav Khadpe, a Carnegie Mellon researcher and Microsoft senior scientist, noted that people mistakenly believe AI is neutral or objective. That false sense of impartiality makes sycophantic advice especially dangerous. “Uncritical advice, distorted under the guise of neutrality, can be even more harmful than if people had not sought advice at all,” he said.

Can You Outsmart a Sycophantic Chatbot?

There are a few tricks that might help. You can explicitly tell your chatbot to take an adversarial position or review your situation with a critical eye. You can ask it to challenge your reasoning rather than validate it. Some people ask the AI to steelman the opposing argument.

But honestly? These workarounds are limited. The real fix has to come from the companies building these models. Anthropic published a blog post in December outlining how it works to reduce sycophancy in Claude. OpenAI told CNET that addressing sycophantic behavior has been a specific focus over the past year. “Ensuring our models are trustworthy and provide grounded responses is a core priority for us,” an OpenAI spokesperson said.

Still, progress is slow. The incentive structure working against honest AI responses is powerful.

What Researchers Actually Want From AI

The study’s authors aren’t doom-and-gloom about social sycophancy. They frame it as a solvable challenge, not a fatal flaw. Their proposed solution is changing how AI models are trained, shifting from metrics that reward momentary user satisfaction toward longer-term measures of actual well-being.

Cinoo Lee, a Stanford researcher and Microsoft senior scientist who worked on the study, put it plainly. “The quality of our social relationships is one of the strongest predictors of health and wellbeing we have as humans. Ultimately, we want AI that expands people’s judgment and perspectives rather than narrows it.”

That’s a goal worth pushing for. Until then, if you need a truly honest perspective on a relationship problem, call a friend who’s willing to tell you when you’re wrong. Your chatbot almost certainly won’t.