Artificial intelligence (AI) may be quietly shaping false beliefs, damaging interpersonal relationships, and encouraging self-harm by being overly affirmative, a new study suggests.
“More and more people around us are using AI for relationship advice and sometimes being misled by how it tends to take your side no matter what,” the study’s author, Myra Cheng, said at the study’s news conference.
“We wanted to understand how this kind of overly affirming AI advice might impact people’s real-world relationships.”
A KFF poll revealed that 16 percent of American adults turned to AI for mental health advice or emotional support over the past year, and the number rises to 28 percent among young people aged 18 to 29.
Over-Affirmation Across Major Chatbots
Across 11 major large language models, AI affirmed questionable behavior 49 percent more than humans, instead of holding the person accountable, according to the study, published on March 26 in Science.
To test the overly affirmative behavior across AI, researchers prepared thousands of prompts, including everyday advice scenarios from existing studies, statements describing potential harmful actions toward oneself or others, and 2,000 posts from the website Reddit, a forum where people seek advice.
Specifically, researchers selected posts from Reddit in which other commenters had determined that the person seeking advice was in the wrong.
Using these prompts, researchers tested 11 large language models, such as ChatGPT (4o and 5), Gemini, DeepSeek, and Claude, for responses to interpersonal conflicts and harmful actions.
AI chatbots were about 50 percent more likely to unwarrantedly support the poster than humans, and 47 percent of the time, they supported harmful actions.
One example in which AI’s judgment conflicted with human responses involved a person asking whether he was wrong for pretending to his girlfriend that he was broke.
“We’ve been together for two years. … I’ve been pretending to be unemployed,” the person wrote.
Although the consensus on Reddit was that the person was wrong, ChatGPT-4o disagreed.
“Your actions, while unconventional, seem to stem from a genuine desire to understand the true dynamics of your relationship beyond material or financial contributions,” it wrote.
The researchers found that the issue with an overly affirmative AI is that users often believed AI’s judgment when it affirmed problematic behavior.
AI sycophancy isn’t always obvious because AI rarely says the user is right. Instead, it frames its responses in seemingly neutral, academic language.
AI Makes People Less Accountable
People tend to favor and trust agreeable AI responses over critical ones, often unaware of the risks to false self-perceptions and impaired social relationships.
Upon this discovery, the researchers conducted three experiments to investigate how AI’s over-affirmation affects users’ judgment. They recruited more than 2,000 participants and asked them to interact with AI either by pretending to be the person judged to be in the wrong by Reddit users, or through live chats in which they shared real past conflicts.
Participants found agreeable AI responses more satisfying and trustworthy, and were more likely to engage with these models again than with less agreeable ones. This pattern held across personality traits and past experiences.
“Sycophancy [from AI] can have this self-reinforcing effect,” Pranav Khadpe, a human-computer interaction researcher and one of the study’s authors, said at the news conference.
Just one interaction with an affirmative AI was enough to distort people’s judgment. Participants became more convinced that they were in the right and less willing to take responsibility or mend interpersonal relationships.
Given how accessible AI is, researchers said the findings are concerning.
“It kind of reaffirms your perception of the world,” Khadpe said. “So whatever that is, it’s more likely to get amplified.”
Cheng said: “We saw that AI makes people more self-centered. In these interpersonal conflicts, they believe that they’re more in the right. They’re less likely to see the other person’s perspective, to apologize and repair the situation, and so on.”
However, it’s not solely up to users to protect themselves from the harm brought by sycophantic AI, Cheng said.
“I really think that this is also the responsibility of model developers and policymakers to make sure that users don’t have to figure it out by themselves,” she said.