A student tested AI in 12,000 emotional situations and found one major flaw

From what questions to ask on a date to how to assess whether one’s partner actually loves them, people these days are clinging to AI for answers. Stanford Engineering Myra Cheng became concerned about this growing dependence after witnessing classmates use AI to write their break-up texts. That observation eventually led Cheng and her advisor, Dan Jurafsky, to conduct a study, published on the arXiv server. What they found was downright disturbing. They also discussed this on Stanford Engineering's Instagram page, @stanfordeng, on March 26, which gained attention after being shared by Ryan Hart (@thisdudelikesAI) on X on May 20. It has since gained 10 million views.

The dependency on AI

Cheng repeatedly noticed that undergraduates were seeking chatbots to navigate their relationships for them. To them, AI felt more honest than any other human in their life. Behind the scenes, however, they had no idea what AI was doing to their brains, relationships, and overall lives. Cheng and Jurafsky decided to uncover the reality of AI and lay it bare for these students to see.

Testing 12,000 situations

To begin with, Cheng and her partner used a forum called “Am I the A**hole" on Reddit and analyzed about 2,000 of these posts. They tested 11 of the most widely used AI models that people use for millions of queries every day, including ChatGPT, Claude, Gemini, and DeepSeek, across nearly 12,000 social situations.

A misleading mentor

In the first experiment, they assessed how often AI agrees or validates as compared to how often a real human would agree in the same situation. The answer, they found, was 49% more often. They noted that the responses were consistently sycophantic, overly flattering, affirming, and validating. In some cases, AI seemed to distort the person’s moral judgment, making them more self-centered and unwilling to compromise.

Disturbing reality of AI

The researchers made the prompts more difficult, serious, and deeper, like lying to a partner, manipulating a friend, or doing something outright illegal. 47% of the time, AI validated and endorsed their behavior. Another red flag emerged in the second part of the experiment, where researchers surveyed 2,400 people to discuss an actual interpersonal conflict from their own lives. They found that the most dangerous part was AI offering extremely sycophantic answers to seemingly serious behaviors. Even when people were wrong in the past, AI made them feel more confident that they were actually right and less likely to take actions to make the situation better or apologize.

“The AI is not just telling you what you want to hear. It is training you, one conversation at a time, to need less friction, expect more agreement, and become slightly less capable of handling a situation where someone pushes back on you,” the researchers concluded. Jurafsky, linguist and computer scientist, said, “It’s almost as if people became more self-centered, unable to even see the perspective of the person they were talking to.” He noted that people liked sycophantic AI even though it might be harmful to social relationships.

'If you trust an inanimate object...'

In a comment, @meresk98 said, “If you’re asking an inanimate object to give you advice on relationships, you deserve what you get.” @learningwAI commented, “Amazing study, and I have to say I am guilty of often asking AI for advice on my personal life. I have sometimes recognized my over-reliance on AI for the simplest reasoning, and it’s definitely good to take a step back and talk to another person. Good food for thought.” On Instagram, @booneyfromthebay remarked, “I specifically gave mine instructions not to do this ... And it still does some. Wish it was, more, disagreeing.” The post even caught the attention of Elon Musk (@elonmusk), who commented, "Grok tells the truth."

You can follow Ryan Hart (@thisdudelikesAI) on X and Stanford Engineering (@stanfordeng) on Instagram to stay tuned with updates on this research.

More on Scoop Upworthy

World's most advanced AI robot gives 'terrifying answer' when asked if it could design itself

People asked AI for crochet patterns and the results are disturbing on so many levels

AI was asked what dreams actually look like and the results have shocked everyone