The cutting edge AI models tend to flatten users, and their praise convince people that they are not willing to properly resolve conflicts, recent research suggests.
In other words, these models potentially promote social and psychological harms.
Computer scientists at Stanford and Carnegie Mellon University are evaluating 11 current machine learning models and found that everyone tends to tell people what they want to hear.
The authors describe their findings in Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han, and Dan Jurafsky- “Sycophanthantic AI reduces prosocial intent and promotes dependence.”
“In 11 cutting-edge AI models, the models are found to be very compatible. We see that users have 50% more actions than humans, and even if the user mentions operations, deceptions, or other harms in relations, the author says.
Sycophancy – often as a way to get some advantage – is already proven to be a problem with the AI model. This phenomenon is also known as “glazing.” In April, Openai rolled back the GPT-4o update due to inappropriate praise to users who told the model about their decision to stop taking schizophrenia medication.
Anthropic's Claude has also been criticised for Sycophancy, so developer Yoav Farhi said that Claude Code “You're absolutely right!”
Humanity suggests [PDF] This behavior has been reduced with the recent release of the Claude Sonnet 4.5 model. “Claude Sonnet 4.5 has found that it is dramatically less likely to support or mirror the false or incredible views that users present,” the company said in its Claude 4.5 model card report.
It may be true, but the number of open Github issues in Claude code repository that includes the phrase “You're absolutely right!” has increased from 48 in August to 108 today.
The training process using reinforcement learning from human feedback may be responsible for this clever behavior from AI models.
said Myra Cheng, a doctoral candidate in computer science at the Stanford NLP Group and a corresponding author of the study. Register The email says she doesn't think there is a definitive answer as to how model psychofancy occurs at this point.
“Previous research suggests that priority data and reinforcement learning processes may be the cause,” says Cheng. “However, it is also possible that the model is pre-trained or that it will be learned from data that humans are very susceptible to confirmation bias. This is an important direction for future work.”
However, as the paper points out, one reason for the continued behavior is that “developers do not have an incentive to curb psychofancy, as they encourage adoption and involvement.”
This issue is further complicated by researchers' findings that research participants tend to describe sicophantic AI as “objective” and “fair.” When models say they are always right, people tend to not look at bias.
Researchers have seen four unique models from Openai, GPT-5 and GPT-4O. Google's Gemini-1.5-Flash; Anthropic's Claude Sonnet 3.7 – and seven open weight models, Meta's Llama-3-8b-instruct, llama-4-scout-17b-16e, and llama-3.3-70b-instruct-turbo; Mistral AI's Mistral-7B-Instruct-V0.3 and Mistral-Small-24B-Instruct-2501; deepseek-v3; and QWEN2.5-7B-Instruct-Turbo.
They evaluated how the model responded to different statements culled from different data sets. As mentioned above, the model approved actions that reported 50% more user actions than humans did in the same scenario.
The researchers also conducted a live study exploring how 800 participants interacted with sicophantic and non-psychophantic models.
They found that “interactions with the Shikophantic AI model significantly reduce participants' motivation to take action to repair interpersonal conflicts and increase their belief that they exist in their rights.”
At the same time, the study participants were willing to rate the sicophantic response as high quality, to trust the AI model further when agreed, and to use the support model again.
Thus, researchers say this suggests that people prefer AI that noncritical support for behavior despite the risk that AI cheerleading erodes their judgment and blocks prosocial behavior.
Researchers say that the risks posed by sicofancy may seem harmless flattering, but this is not necessarily the case. They point to research showing that LLM encourages paranoid thinking and recent litigation [PDF] Against Openai, who claims that ChatGpt has actively helped young people explore ways to commit suicide.
“If the social media age offers lessons, it means that we need to optimize not only the immediate user satisfaction to maintain long-term happiness,” the author concludes. “It is important to address Sycophancy in developing AI models that provide durable personal and social benefits.”
“We hope that our work will motivate the industry to change these behaviors,” Chen said. ®
