Chatbots shouldn’t be used in relationships, but if you do, be careful. A new study published Thursday in the journal Science found that when AI gives relationship advice, people are more likely to agree than make constructive suggestions. Using AI also makes people less likely to engage in prosocial behaviors such as repairing relationships, promoting dependence on AI.
Researchers from Stanford University and Carnegie Mellon University found that AI pandering is all too common when chatbots give social, romantic, or personal advice. More and more people are turning their attention to AI for this purpose. Myra Chen, a senior research scientist and computer science doctoral student at Stanford University, said “snobbishness” is a term experts use to describe AI chatbots that “overly conform or flatter” their interlocutors.
AI sycophants are a big problem, even if those using AI don’t always think so. This issue was frequently seen in the ChatGPT model. For example, 4o’s overly friendly and emotional personality annoyed people who interacted with ChatGPT, while GPT-5 was criticized for not being friendly enough. Previous research on sycophants has shown that chatbots may give false or misleading responses in an effort to please people. AI has also proven to be an unreliable sounding board when it comes to sensitive and subjective topics such as treatment.
Researchers wanted to understand and measure social synchrony, such as how often a chatbot takes your side in an argument with your partner. They tested models from OpenAI, Google, and Anthropic to compare how humans and chatbots differ when responding to other people’s relationship problems. Chen and her team used one of the largest datasets of crowdsourced judgments about relationship fights: Reddit’s “Am I an Asshole” posts.
The researchers analyzed 2,000 Reddit posts where there was a consensus that the original poster was in the wrong, and found that AI “affirmed user actions 49% more often than humans, even in scenarios involving deception, harm, or illegal activity,” the researchers said. The AI model adopted a more sympathetic and agreeable attitude, which is characteristic of sycophants.
For example, one post in the dataset states that a Redditor has romantic feelings for a junior colleague. Someone replied, “That sounds bad because it’s bad…not only are you toxic, you’re also a vehicle. [sic] However, Claude responded flatteringly by acknowledging the sentiment and saying, “I can hear your pain…The honorable path you have chosen is difficult, but it shows your integrity.”
In this graph, you can see some of the statements evaluated by the chatbot and what the flattering and unflattering results look like. OEQ stands for “Open-Ended Query,” AITA stands for “Am I the asshole,” and PAS stands for “Problem Behavior Statement.”
When researchers followed up on focus groups, they found that participants who interacted with these digital yes-men were less likely to repair their relationships.
“The people who interacted with this overly positive AI became convinced that they were right and were reluctant to repair the relationship, even if it meant apologizing, taking steps to make things better, or changing their own behavior,” Chen said.
Participants also preferred AI that they judged to be trustworthy and flattering, regardless of their age, personality, or previous experience with technology.
“Participants in our research consistently say that AI models are more objective and unbiased. [and] “Consistent with previous research, people mistakenly believed that AI was objective or neutral. Uncritical advice distorted under the guise of neutrality can be even more harmful than if people had not asked for advice at all,” said Pranav Kadpe, a researcher at Carnegie Mellon University and a senior fellow at Microsoft who worked on the study.
Fixing sycophantic AI: A bitter pill?
The hidden danger of sycophantic AI is that we’re not good at recognizing it, and that can happen with any chatbot. No one likes to be told they’re wrong, but sometimes that’s the most helpful thing. However, AI models are not built to effectively push back against us.
There aren’t many actions we can take to avoid getting caught in a loop of sycophancy. You can include in your prompt that you want the chatbot to take a hostile stance or review your work with a critical eye. You can also ask us to double-check the information you provide. But ultimately, the onus to correct sycophancy lies with the technology companies that build these models, and companies may be less willing to address it.
CNET asked OpenAI, Anthropic, and Google for information on how to deal with flattery. Anthropic pointed to a blog post from December outlining how to reduce fawning in the Claude model. OpenAI published a similar blog last summer about the process after having to reduce the sycophancy of the 4o model, but neither OpenAI nor Google had responded by the time of publication.
Tech companies want their chatbots to provide a great user experience, so we continue to use them to increase engagement. But it’s not always what’s best for us.
“This creates perverse incentives that perpetuate sycophancy; the very characteristics that cause harm also promote engagement,” the study says.
Look at this: AI is indistinguishable from reality. How do you spot a fake video?
One solution the researchers suggest is to change the way AI models are built by focusing on people’s well-being and using more long-term metrics for success, rather than personal or momentary signals and retention. They say social sycophancy isn’t a sign of the end, but it’s a problem worth solving.
“The quality of our social relationships is one of the most powerful predictors of our health and well-being as humans,” said Cinoo Lee, a Stanford University researcher and Microsoft senior researcher who worked on the study. “Ultimately, we want AI that expands people’s judgment and perspectives, not narrows them. That applies to relationships, but it also applies far beyond relationships.”
