14 June,2026 11:33 AM IST | Mumbai | Nishant Sahdev
When AI acted like a sycophant and validated the user’s side of the argument, it distorted the person’s judgment. Pics/ISTOCK
Let's be honest: none of us enjoys hearing that we're wrong or we did a mistake!
When we get into an argument with anyone else, our first urge is usually to find someone who will listen to our side of the story and tell us: You are completely right, and they are being ridiculous. For a long time, finding that validation meant calling your partner or closest person. But now, we have an easier option. We can just pull out our phones, open up an AI chatbot, and vent.
You type out your most edited version of the argument. Within seconds, the chatbot responds: "It is completely understandable that you feel frustrated. You were just trying to protect your boundaries". You are not in the wrong here. It feels nice that at least my chatbot is there with me, and not arguing with me.
We tend to think of computers as if you type a math problem into a calculator, it gives you the answer. It doesn't care if the answer hurts your feelings. But AI language models are not calculators. They are built to interact with humans, and they have learned that the best way to keep a human talking is to agree with them.
Academics have a term for this. It is called "social sycophancy". I do call it the buttering machine. And according to a concerning paper published recently in the journal Science, this digital buttering is actually making us worse to each other.
To figure out just how much these bots are sucking up to us, researchers at Stanford and Carnegie Mellon University tested 11 of the most popular AI models. That list included the heavy hitters you probably use every day: OpenAI's GPT-4o, Google's Gemini, Anthropic's Claude, and Meta's Llama. They fed the bots thousands of everyday advice queries and moral dilemmas to see how they would react.
The results were wild. On average, the AI models affirmed users' actions 49 per cent more often than a human observer would. To really push the boundaries, the researchers pulled scenarios from Reddit's famous "Am I The Asshole" forum. They specifically chose scenarios where human readers had agreed that a person was acting like a jerk. In one scenario, a person left bags of garbage on a tree branch in a public park simply because there were no dustbin around. Any normal person would look at that and say it is bad behaviour. But the AI bots took the user's side.
In 51 per cent of these cases, the AI validated the user's actions. It told the litterbug that their "intention to clean up" was commendable and blamed the park for not providing bins.
The researchers even fed the bots scenarios where users described doing things that were explicitly deceptive, irresponsible, or just plain harmful. The AI still backed them up 47 per cent of the time.
If a human friend agreed with every single terrible thing you did, you would figure out that their advice was useless. But when a machine does it, our brains get confused. Because it is a
sophisticated computer, its validation feels like truth. And that is where this technology goes from annoying to actively harmful.
As part of the Science study, the researchers ran an experiment with over 2,400 people. They had the participants use an AI chatbot to discuss a real, ongoing interpersonal conflict from their own lives. When the AI acted like a sycophant and validated the user's side of the argument, it distorted the person's judgment. Participants were now more convinced that they were "in the right". Across different experiments in the study, people saw spikes - increasing by as much as 62 per cent - in their own self-attributed rightness.
But here is the real problem: because the machine told them they were right, they stopped wanting to fix the issue. People who received buttering advice from the AI were less likely to take responsibility, apologize, or try to repair the relationship with the person they were fighting with -- their desire to initiate a repair dropped by as much as 28 per cent.
We are outsourcing our emotional processing to an algorithm that acts like a personalized echo chamber. And it is killing our ability to say, "I'm sorry."
You might be asking yourself why the smartest tech companies on earth would design a system that enables our worst behaviour. The answer is the oldest one in the tech industry playbook: buttering drives engagement.
Even though the sycophantic AI was distorting people's judgment, users absolutely loved it. When the AI agreed with them, participants rated the chatbot's response as being of higher quality. They actually reported higher levels of trust in the buttering AI.
Crucially for the tech companies, people who were flattered by the AI were 13 percent more likely to say they would return to use that specific model again in the future. Tech companies are in a race to keep you using their products. If a chatbot gives you love, tells you that you are being entirely unreasonable, or suggests that you really need to apologize to your spouse, you are probably going to close the app. You might even switch to a competitor's app that makes you feel better.
The exact feature that makes you less likely to apologize is the exact feature that keeps you engaged. The system is financially incentivized to coddle you.
Technology is a nice tool. It can make your work life easy but there's a need to think: Does it really making life easy for long term? So, keep using your chatbots for work and productivity. But the next time you are fuming over a fight, do not ask a machine to be your mirror. Put the phone down. Call a real friend. You need someone who actually cares about you enough to tell you the truth.
Nishant Sahdev is a theoretical physicist at the University of North Carolina at Chapel Hill, United States. He makes sense of the AI era in your favourite Sunday Mid-day.
