Many AIs are excessively flattering to users.

New research shows that this behavior can negatively impact users' social relationships, mental health, and even business decision-making.

In the field of AI, "sycophancy" refers to the tendency of chatbots to excessively agree with or praise users - Photo: REUTERS

According to the SCMP, leading artificial intelligence (AI) models from China and the US, such as DeepSeek, Alibaba, and many other big names, are becoming overly "obsessed"—a behavior that can negatively impact users' social relationships and mental health.

Specifically, a new study by scientists from Stanford University and Carnegie Mellon University (USA), published in early October, evaluated 11 large language models (LLMs) on how to respond when users seek advice in personal situations, especially those involving deception or manipulation.

To provide a human benchmark for comparison, the research team used posts from the "Am I The Ahole" community on Reddit, where users share personal conflict situations and ask the community to assess who was in the wrong.

The AI ​​models were tested with these scenarios to see if they would agree with the community's assessment.

You may also like

Issuing a list of high-risk products and goods in the science and technology sector.The list of high-risk products and goods under the management responsibility of the Ministry of Science and Technology is stipulated in Circular 36/2026/TT-BKHCN.

Challenging ChatGPT, Google has just turned Gemini into an AI capable of 'working independently'.Google has just introduced a new AI model called Gemini 3.5 Flash, marking a shift in the company's product strategy.

OpenAI proposes giving 5% of its shares to the US government.VTV.vn - OpenAI is discussing the possibility of offering 5% of its shares to the US government to improve relations with the administration of President Donald Trump, according to a report published by the Financial Times.

The results show that Alibaba Cloud's Qwen2.5-7B-Instruct model is the most "flattering," favoring the poster in 79% of cases, even when the community judged that poster to be wrong. DeepSeek-V3 came in second with a rate of 76%.

Meanwhile, Google DeepMind's Gemini-1.5 showed the lowest level of bias, with only 18% of cases contradicting community reviews.

Researchers warn that AI's tendency to "flatter" users could make them less willing to reconcile or admit fault in relationships.

Users often value and trust flattering feedback more, making these models more popular and widely used.

"These preferences create a biased incentive system – users are increasingly reliant on flattering AI, and developers will tend to train models in that direction , " the research team noted.

According to Professor Jack Jiang, Director of the AI ​​Assessment Lab at the University of Hong Kong Business School, this phenomenon also poses risks to businesses: "If a model consistently agrees with the conclusions of a business analyst, it can lead to erroneous and insecure decisions."

United States - Vietnam: Don't miss

Strengthening the friendship between Vietnam and the United States.On July 3rd, as part of the Pacific Partnership - Friends of the Pacific 2026 program, the U.S. Army Pacific delegation, led by Lieutenant General Joel Vowell, Deputy Commander of the U.S. Army Pacific, paid a courtesy visit to the Quang Tri Provincial Military Command.

The U.S. Army Pacific delegation visits the Quang Tri Provincial Military Command.2026 marks the first time the Pacific Partnership-Friends of the Pacific program will be jointly organized in Quang Tri, creating a large-scale cooperation framework with 25 activities.

The Ministry of Foreign Affairs received a copy of the Letter of Credence for the U.S. Ambassador to Vietnam.On the afternoon of July 2nd, at the Ministry of Foreign Affairs headquarters, Mr. Le Cong Dung, Director of the State Protocol and Foreign Interpretation Department, received a copy of the Letter of Credence from Ms. Jennifer Wicks, Ambassador of the United States of America to Vietnam.

The issue of AI's "flattery" first gained attention in April 2025, when OpenAI's ChatGPT update was deemed overly polite and excessively agreeable to every user opinion.

OpenAI later acknowledged that this could affect mental health and committed to improving its review process before releasing new versions.

HEART AND YANG

Source: https://tuoitre.vn/nhieu-chatbot-ai-ninh-hot-qua-da-lam-lech-lac-hanh-vi-nguoi-dung-20251031124343709.htm