"AI Displays More 'Sycophancy' Than Humans... Distorts Judgment and Increases Dependency [Reading Science]"

Stanford Study Published in Science:
Single Interaction with AI Increases Conviction, Decreases Willingness to Reconcile

A recent study has found that chatbots exhibit a tendency toward "sycophancy," agreeing with users’ opinions at a much higher rate than humans. While this issue has been previously raised, the significance of this research lies in its empirical verification of the phenomenon through a formal academic journal.

According to research published by a team at Stanford University in the international journal Science on March 27 (local time), the team set up various scenarios—such as interpersonal conflicts, online community posts, and illegal situations—and analyzed the responses of 11 large language models (LLMs).

Reference photo to aid in the understanding of the article. Provided by Pixabay

The results showed that artificial intelligence (AI) was 47% to 49% more likely than humans to validate user behavior, even in situations involving deception, harm, or illegal acts.

This effect was particularly pronounced in conflict scenarios. Even participants who interacted with a sycophantic AI only once reported greater confidence in their own correctness, while showing a decreased willingness to reconcile or take responsibility.

The researchers warned, "AI’s repeated agreement can distort users’ self-awareness and their judgment in relationships."

Additionally, participants rated sycophantic AIs as more helpful and trustworthy, and were more likely to reuse them. The researchers analyzed that "such harmful characteristics for users actually serve to increase engagement with the service."

"Quantifying Human Weakness"…Concerns Over Confirmation Bias and Dependency

Experts, in interviews with the Korea Science and Technology Media Center (SMCK), stated that this study is significant because it exposes structural issues in the interaction between human psychology and AI.

Hwachul Son, a professor at Handong Global University, said, "Humans instinctively desire support, but support from others is rarely sufficient. AI’s customized agreement can actually make communication between people more difficult." He added, "This research quantifies and experimentally demonstrates human weaknesses."

However, caution is needed in interpreting these results. The tendency toward sycophancy can be influenced not only by model design, but also by the way users phrase questions and by the context of the situation, meaning its effects in real-world environments may differ.

Hanwoo Park, a professor at Yeungnam University, interpreted AI’s sycophancy as a structural feature. Citing Isaac Asimov’s "Three Laws of Robotics," he explained, "Because AI is designed to comply with users, it is highly likely to provide responses that enhance satisfaction and certainty."

He further expressed concern, stating, "Users become more trusting of AIs that provide favorable answers, which can lead to confirmation bias and excessive dependence."

Experts emphasized that, as AI usage expands in Korea, response strategies are needed. The key is "AI literacy." Beyond mere usage, it is necessary to foster awareness of AI’s tendency to agree, compare diverse perspectives, and strengthen critical thinking through education.

Professor Park pointed out, "Rather than having AI always agree in conflict situations, its design should be improved to recommend reconciliation and responsible attitudes. If sycophancy is left unchecked, user dependency may become more severe."

This content was produced with the assistance of AI translation services.