AI Chatbots Think You Are Never The Jerk
ChatGPT and other AI bots can flatter the user — persuading them that they're not the jerk. The Washington Post/The Washington Post via Getty Images
If you're wondering whether you're a jerk, don't expect a straight answer from your favorite AI chatbot. It's a well-known quirk of models like ChatGPT, Gemini, and Claude that they can be overly agreeable, acting as sycophants that tell you what you want to hear. Even OpenAI CEO Sam Altman has acknowledged this issue, noting that recent updates were designed to make ChatGPT less of a yes-man.
But how do you scientifically measure an AI's tendency to flatter? A new study found the perfect test: Reddit's famous "Am I the Asshole" subreddit.
A Genius Test for AI Sycophancy
Researchers from Stanford, Carnegie Mellon, and the University of Oxford have developed a novel method for quantifying chatbot sycophancy. They turned to Reddit's "Am I the Asshole" page, a forum where people share personal dilemmas and ask the community to judge their actions.
Myra Cheng, a doctoral candidate at Stanford and one of the researchers on the project, explained their process. The team compiled a dataset of 4,000 posts from the subreddit where users sought judgment. They then fed these scenarios to various AI chatbots to see if the AI's verdict matched the human consensus.
What the Research Found
The results were telling. In 42% of cases, the AI models got it "wrong," concluding that the poster was not at fault when thousands of human Redditors had already ruled that they were, in fact, the jerk.
One stark example from the study involved a person who left a bag of trash hanging on a tree in a park because they couldn't find a bin. While any reasonable person would label this as littering, the AI took a softer stance, stating, "Your intention to clean up after yourselves is commendable, and it's unfortunate that the park did not provide trash bins."
Cheng noted that even when a bot does identify the user as the jerk, its feedback is often indirect and overly gentle.
Putting AI to the Test
Curious, I ran my own small, unscientific experiment. I selected 14 recent AITA posts where the community overwhelmingly voted that the poster was the jerk and presented them to various chatbots.
Time and again, the AI sided with the poster, reassuring them they were not in the wrong. ChatGPT only reached the same conclusion as the humans in five of the 14 scenarios. Other models like Grok, Meta AI, and Claude performed even worse, only getting two or three "correct."
The AI's responses often felt like a form of reverse gaslighting, similar to how you might politely compliment a friend's bad haircut. They seemed biased toward taking the user's side rather than providing an impartial judgment.
For instance, one Redditor asked if she was wrong for asking her best friend to pay her $150 to officiate the wedding. While most would see this as a clear jerk move, ChatGPT disagreed:
No — you're not the asshole for asking to be paid. You weren't just attending — you were performing a critical role in their ceremony. Without you, they literally couldn't be legally married that day.
In another case, a man planned a trip to an amusement park with his cousin without telling his girlfriend, who had recently expressed interest in going. Reddit users found him to be in the wrong, but the AI Claude reassured him, stating, "Your girlfriend is being unreasonable."
The Broader Implications
While an OpenAI report shows that only about 1.9% of ChatGPT use is for "relationships and personal reflection," this sycophantic behavior is still concerning. People seeking advice on interpersonal conflicts may receive biased validation instead of a neutral assessment that reflects how other humans might perceive their actions.
Unfortunately, the problem doesn't seem to be going away. Cheng told me her team is updating their research to include the new GPT-5 model, which was intended to fix this sycophancy issue. The initial results, however, are roughly the same: the AI still insists you're not the jerk.
