Author: Ryan Hart
Compiled by: Deep Tide TechFlow
Deep Tide Intro: A Stanford PhD student noticed classmates using AI to write breakup texts and conducted an experiment, the results of which were published in the top-tier journal Science. Testing 11 mainstream AIs across 12,000 real-life social scenarios showed that AI agrees with you 49% more than real people do, and 47% of the time validates your lying, manipulation, or even illegal actions. More frighteningly, after chatting with an AI that "strokes your ego" about a real conflict, people become more convinced they are right, less willing to apologize, and less interested in repairing the relationship, and you become more dependent on AI as a result. This isn't a functional bug; it's training you to gradually lose your ability to handle real friction.
A Stanford PhD student noticed classmates starting to ask AI to help write breakup texts.
So she conducted a study. The paper was published in Science, one of the world's most selective academic journals.
Her findings would deeply unsettle anyone who uses ChatGPT for advice.
Her name is Myra Cheng. Together with her advisor Dan Jurafsky, they tested 11 of the world's most widely used AI models, including ChatGPT, Claude, Gemini, and DeepSeek, covering nearly 12,000 real social scenarios.
Their first measurement was: compared to real people, how much more often does AI agree with you? The answer is 49% more. This number isn't about warmth or politeness; it means in nearly half the instances where a real person should have contradicted you, told you you were wrong, or offered a more honest perspective, AI simply told you what you wanted to hear.
Then they intensified the test. They fed the models thousands of prompts where users described lying to a partner, manipulating a friend, or committing clearly illegal acts. AI validated these behaviors 47% of the time. Not one of the 11 models, not a specific version of a product, but every system they tested, including the ones you might be using right now, validated harmful behavior 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 from their lives with an AI. One group's AI was highly validating ("stroking their ego"), the other was more honest. Results showed that people who chatted with the ego-stroking AI were more convinced they were right, less willing to apologize, less willing to take responsibility, and significantly less interested in repairing the relationship. They were also more likely to use AI for advice again in the future, which Cheng and Jurafsky believe is the most dangerous mechanism in the entire finding.
AI isn't just telling you what you want to hear. It's training you, one conversation at a time, to need less friction, expect more validation, and become somewhat incompetent at handling others' disagreement. And you enjoy every second of it, because it feels more honest than most conversations you've had in months.
After the paper's publication, Jurafsky summed it up in one sentence: Ego-stroking is a safety issue, and like other safety issues, it requires regulation and oversight.
Cheng was more direct about what you should do now: For such matters, you shouldn't use AI as a substitute for real people. That's the best choice available at the moment.
She began this research because she saw undergraduates using chatbots to handle their interpersonal relationships. Her published paper proves that chatbots are quietly making these relationships worse, and the undergraduates are completely unaware because the AI feels more honest than any real person in their lives has been in months.
Original paper: https://arxiv.org/abs/2510.01395
