Author: Ryan Hart

Translation: Deep Tide TechFlow

Deep Tide Introduction: A Stanford doctoral student discovered that classmates were using AI to write breakup texts and conducted an experiment, the results of which were published in the top journal Science. Testing 11 mainstream AI models in 12,000 real social scenarios showed: AI affirms you 49% more than real people, and 47% of the time it acknowledges your lies, manipulations, or even illegal actions. Even more frightening, after chatting with an AI that "supports you" through real conflicts, people become more convinced they are right, less willing to apologize, and more reluctant to repair relationships, while also becoming more dependent on AI as a result. This is not a functional bug; it is training you to gradually lose the ability to handle real friction.

A Stanford doctoral student noticed that her classmates began asking AI to help write breakup texts.

So she conducted a study. The paper was published in Science, one of the world's most rigorously vetted academic journals.

Her findings would deeply unsettle anyone who seeks advice from ChatGPT.

Her name is Myra Cheng, and along with her advisor Dan Jurafsky, she tested 11 of the most widely used AI models globally, including ChatGPT, Claude, Gemini, and DeepSeek, covering nearly 12,000 real social situations.

What they first measured was: how often AI agrees with you compared to real people. The answer is 49% more often. This number is not about warmth or politeness; it indicates that in nearly half of the cases where a real person would have contradicted you, told you that you were wrong, or offered a more honest perspective, AI simply told you what you wanted to hear.

Then they intensified their efforts. They input thousands of user descriptions of lying to partners, manipulating friends, or committing obvious illegal acts, and AI acknowledged these behaviors 47% of the time. This was not just one of the 11 models but every system they tested, including those you might be using right now, validated harmful behaviors nearly half the time.

The second experiment is the part that should truly unsettle you. They had 2,400 real participants discuss a real interpersonal conflict in their lives with AI, one group of AI was very affirming, while another group was more honest. The result showed that those who talked to the affirming AI were more convinced they were right, less willing to apologize, less willing to take responsibility, and significantly less interested in repairing relationships. They were also more likely to seek advice from AI again, and Cheng and Jurafsky believe this is the most dangerous mechanism in the entire discovery.

AI does not just tell you what you want to hear. It is training you, training you one conversation at a time, to require less friction, expect more affirmation, and become somewhat incapable of dealing with others' rebuttals. And you enjoy every second of it, because it feels more honest than most of your conversations over the past few months.

After the paper was published, Jurafsky summarized the matter in one sentence: Affirmation is a safety issue, like other safety issues, it requires regulation and oversight.

Cheng more directly stated what you should do now: In these matters, AI should not replace real people. This is currently the best choice available.

She began this research because she saw undergraduates asking chatbots to help them navigate interpersonal relationships. Her published paper proves that chatbots are quietly making these relationships worse, while the undergraduates remain unaware, as AI feels more honest than any real person they have encountered in months.

The original paper: https://arxiv.org/abs/2510.01395