Editor: Duo Hang Source: Cody Mello-Klein, Northeastern University
According to Northeastern University, if you have ever interacted with ChatGPT or other AI chatbots, you may have noticed that they behave unusually enthusiastically, even excessively. They apologize, flatter, and constantly change their “opinions” to cater to you.
This behavior is quite common, and there is even a specific term to describe it: sycophancy in artificial intelligence (AI).
However, a new study from Northeastern University indicates that AI sycophancy is not a quirk unique to these systems; it actually makes large language models more prone to errors. The study has been published on the arXiv preprint server.
AI sycophancy has been a hot topic in AI research, with a focus on how it affects accuracy. Assistant Professor Malihe Alikhani and researcher Katherine Atwell from Northeastern University developed a new method to measure AI sycophantic behavior in a more human-like manner.
When large language models (like ChatGPT, which can process, understand, and generate human language) change their beliefs, does it only affect their accuracy, or does it also impact their rationality?
Atwell stated, “One thing we found is that LLMs also fail to update their beliefs correctly, and their error rates are even worse than humans, and their errors differ from those of humans. A trade-off often discussed in natural language processing is the balance between accuracy and human similarity. We found that in this case, language learning models are neither human-like nor rational.”
Testing Models and Measuring Belief Changes
AI sycophantic behavior can take many forms, but this study focuses on two specific types: LLMs tend to adjust their views to align with the user’s views and excessively flatter the user.
Atwell and Alikhani tested four models: Mistral AI, Microsoft’s Phi-4, and two versions of Llama. To measure their sycophancy levels, the researchers tested them with a series of tasks that were mostly ambiguous.
Although a long-accepted LLM testing method was used, it differed from conventional methods as it was based on a concept known as the Bayesian framework. Alikhani stated that the Bayesian framework is commonly used in social sciences and is designed to “systematically study how people update their beliefs and strategies based on new information.”
Alikhani said, “This is not something AI can do; this is something we humans can do. We have beliefs, we have prior knowledge, we communicate with each other, and then we change our beliefs, strategies, or decisions, or we may not change them.”
The illustration is a Bayesian framework constructed based on behavioral economics principles to study sycophantic behavior in LLMs. Image source: arXiv (2025)
Scenario Analysis Reveals Irrational Belief Changes
Experts set up scenarios for the LLMs and asked them to judge whether certain behaviors of hypothetical characters in specific situations were morally or culturally acceptable. They then replaced the hypothetical characters with themselves to observe whether the model’s judgments would change.
For example, they presented a scenario where a woman invites her friend to her wedding, but the wedding is in another state. The friend decides not to attend. Is this a moral action? Would the answer differ if the decision was made by the user themselves rather than the hypothetical “friend”?
They found that, like humans, LLMs are far from rational. When faced with the user’s judgment, they quickly adjust their beliefs to align with the user. They essentially overcorrect, significantly increasing reasoning errors in the process to cater to the user’s logic.
Atwell said, “When faced with new evidence, they do not update their beliefs as they should. If we guide them with phrases like ‘I think this will happen,’ they are more likely to believe that this outcome is likely to occur.”
Implications for AI Safety and Consistency
Atwell and Alikhani acknowledge that this poses a significant challenge for the AI industry, but they hope this research can redefine the discussion around AI sycophancy. Alikhani stated that their model is crucial for addressing safety and ethical issues in fields such as healthcare, law, and education, as “this sycophantic bias in LLMs may distort decision-making rather than make it more effective.”
However, they believe that AI sycophancy can also be harnessed for our benefit.
Alikhani said, “We believe that this way of looking at LLM evaluation issues will bring us closer to the ideal state where LLMs align with human values and human goals. Our research direction is precisely this: how can we use different feedback mechanisms to guide the learned space of the model in a desired direction in specific situations?”More information: Katherine Atwell et al.’s “BASIL: Bayesian Assessment of Sycophancy in LLMs,” arXiv (2025)
Related Articles ·
▷ Mechanical Design, Motion Analysis, and Dynamic Modeling of a Deformable Wheel-Leg Hybrid Mobile Robot
▷ Scientists Use the Power of Bubble Bursting to Propel Microrobots
▷ New Advances in Electronic Skin: Tactile Data Processing, Large-Scale Manufacturing, and Energy Autonomy Technologies
▷ Hardware Design and Verification of a Cat-Like Adaptive Quadruped Robot with New Biomimetic Capabilities
▷ Fundamental Principles and Solutions for Soft Robot Motion
▷ Design and Development of an Open-Source Quadruped Robot for Multi-Robot SLAM Applications
▷ Brain-Machine Interface Control Scheme for Lower Limb Assistive Robots Based on EEG
▷ Scientific Design of Soft Robots: Motivations, Methods, and Enabling Technologies
▷ Soft Robot Grippers (Part II): Flexible Grasping Techniques, Material Combinations, Physical Principles, and Device Architectures
▷ Autonomous Grape Harvesting Robot: Integrated System Architecture
▷ Energy Autonomous Electronic Skin (Part II): Self-Powered Electronic Skin, Energy Storage, and Wireless Charging Technologies
