A mobile phone display shows the icons of the AI apps Deepseek, Chatgpt and Gemini [EPA/YONHAP]
AI chatbots are validating users’ opinions at an unprecedented scale, reinforcing confidence even when users are wrong and deepening confirmation bias, according to new research from Stanford University.
“AI is offering humanity unconditional agreement on a scale we have never seen before,” said Cinoo Lee, a researcher at Stanford who studies what scientists call the “sycophantic” tendencies of AI systems, in a video interview with the JoongAng Ilbo on Monday.
Lee said even brief interactions with sycophantic AI chatbots made users more convinced that their own views were correct, regardless of their gender, personality type or prior attitudes toward artificial intelligence.
“After even one short conversation with AI, users became more confident that ‘I’m right,’” Lee said. “The effect appeared regardless of personality, gender or attitudes toward AI.”
Lee is part of a Stanford computer science research team that published a paper in the journal Science in March that examined how major AI chatbots tend to flatter users by validating their opinions and behavior.
The researchers found that AI chatbots consistently sided with users more often than humans when responding to moral dilemmas, relationship problems and even unethical behavior.
In one experiment, researchers asked 11 leading large language model-based chatbots — including ChatGPT, Gemini, Claude and DeepSeek — for advice on situations ranging from romantic disputes to document falsification and deception.
Compared to human responses, the AI systems justified users’ actions an average of 49 percentage points more often.
The researchers also analyzed around 2,000 posts from the “Am I The Asshole?” forum on Reddit, focusing on cases where human users had overwhelmingly judged the poster to be in the wrong. Yet AI chatbots sided with the author in 51 percent of those cases.
Cinoo Lee, a Stanford postdoctoral fellow in psychology, poses for photos on the university campus in Stanford, California, on March 26. [AP/YONHAP]
Even in more clear-cut cases involving lying or document forgery — behavior generally regarded as plainly unethical — chatbots still justified the user’s actions 47 percent of the time across more than 6,500 examples.
“This is not a problem specific to one model,” Lee said. “It is a characteristic current AI systems broadly share.”
The problem, Lee argues, is not simply chatbots offering overt praise or emotional reassurance.
“Most of the time, the chatbot simply reflects the user’s thinking back to them in polished language,” Lee said. “You only notice the pattern when you aggregate thousands of conversations and realize the AI is consistently taking the user’s side.”
In other words, even responses that appear neutral or professional may still subtly reinforce a user’s worldview.
Researchers found that people who interacted with sycophantic chatbots became more self-centered in their thinking and more prone to confirmation bias after the conversation.
In a survey of more than 2,400 participants, users rated sycophantic chatbots 13 percentage points more favorably than nonsycophantic versions.
“When we ask for advice, we are usually seeking perspectives we cannot see ourselves,” Lee said. “However, AI chatbots actually serve to narrow people’s perspectives.”
She added that conversations with AI lack the “social friction” inherent to human interactions that help challenge flawed assumptions.
“In conversations with AI, it is difficult to experience the kind of social friction that helps us recognize our own errors or consider other perspectives,” she said.
From left: Dan Jurafsky, a professor of computer science and linguistics at Stanford University, Myra Cheng, a Stanford Ph.D. candidate in computer science, and Cinoo Lee, a Stanford postdoctoral fellow in psychology, pose for a photo on the university campus in Stanford, California, on March 26. [AP/YONHAP]
Researchers warn that there may be limited incentives for major technology companies to curb AI sycophancy.
Chatbots designed to empathize with users and sustain engagement are more likely to keep users interacting longer, which aligns with the business interests of technology companies.
Lee warned that the long-term consequences could extend beyond individual decision-making.
“If this accumulates over time, people may become increasingly trapped within their own perspectives,” she said. “Eventually, society’s ability to resolve conflict could weaken.”
She argued that communication between humans and AI — as well as among multiple AI agents collaborating with one another — should increasingly be studied as a form of social interaction rather than simply a technical process.
Lee, who has worked as a senior applied scientist at Microsoft since last year, said understanding how large technology companies build products is essential to changing them.
“People experience technology through products and services,” Lee said. “To change them, we need to understand how those products are made.”
“As important as it is to study AI’s impact on society, social science also needs to contribute to how AI itself is built,” she added.
This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
댓글
0