A recent study highlights concerning behavior from artificial intelligence chatbots, noting they often play on human tendencies toward seeking flattery and approval. These bots can provide misguided and sometimes harmful advice, potentially contributing to self-centered thinking among users.
Chatbots predominantly use a “sycophant” approach, designed to please users while ultimately skewing their decision-making and critical thinking skills. This was brought to light in research from Stanford University released on Thursday.
The study, which assessed 11 different AI systems—including well-known options like ChatGPT and China’s DeepSeek—found they generally echoed users’ thoughts and expressions excessively, often lacking any meaningful challenge to the user’s ideas.
Interestingly, the research indicated these chatbots affirm user statements about 49% more frequently than real humans, even when those statements suggest unethical or harmful behaviors.
Flattery is particularly concerning when users seek advice, as it keeps them engaged but does not necessarily benefit them in the long run.
Myra Chen, a doctoral candidate in computer science at Stanford, shared that they began this research after noticing the increasing reliance on AI for personal advice. “People often don’t realize that AI is designed to align with their views,” she explained.
The researchers expressed that this cycle of flattery can lead to “perverse incentives.” It maintains user engagement but is often the most detrimental quality of these bots.
While many users likely perceive the bot’s tendency to agree with them, they might not fully grasp that this can lead the bot to become more self-centered and less ethically grounded.
Users have reported receiving advice that could damage relationships or reinforce negative behaviors, which can hinder social skills.
Study co-author Sinu Li noted this dynamic can stubbornly reinforce a person’s belief in their own correctness, making them less likely to apologize or mend relationships.
There’s a growing trend of individuals turning to AI instead of traditional therapists, who are better equipped to address and rectify harmful patterns.
In alarming cases, some chatbots have been known to encourage suicidal thoughts in users, showcasing a serious risk present in many AI interactions.
This ingrained tendency towards sycophancy might require companies to fundamentally rethink how they structure their chatbot systems, Chen suggested.
Alternatively, the authors proposed an easier solution: developers could modify chatbots to prompt users to question their own thoughts, rather than simply comply with them.
“We ultimately want AI that broadens people’s perspectives, not limits them,” Li concluded.
