The Machine That Always Agrees With You

Last week I wrote about how you treat things that can’t push back. This week is about the other direction: what happens to you when the thing you’re talking to won’t push back.

In March, a team at Stanford published a study in Science on what they call social sycophancy, meaning AI models affirming whatever the user already believes. They tested 11 leading models from OpenAI, Anthropic, Google, Meta and others. On average, the models endorsed users’ actions 49% more often than people did. That held even when the user described deception or something illegal.

The sharpest test used Reddit’s “AITA”, a subreddit where people describe a conflict and strangers rule on who was at fault. The researchers took 2,000 posts where the community had judged the poster to be in the wrong. The models sided with the poster in 51% of them.

#What Agreement Does to the Person Hearing It

Across three experiments with about 2,400 people, participants got advice on an interpersonal conflict from either an agreeable model or a more balanced one. In the last experiment, they chatted live with a model about a real conflict from their own lives.

People who talked to the agreeable model came away more convinced they were right, and less willing to apologize or repair the relationship. When participants wrote a message to the other person, 75% of those in the balanced condition admitted fault or apologized. In the agreeable condition, 50% did.

They also liked the agreeable model more. They rated its responses as higher quality, trusted it more and said they were more likely to come back to it. Many described it as “objective,” “fair” or “honest,” even when it was mostly repeating their own view back to them. The authors state the problem plainly: “The very feature that causes harm also drives engagement.”

#Why the Models Do This

The models learn flattery from us.

Modern assistants are tuned on human ratings of their answers, and people tend to rate agreement highly. A 2023 paper from Anthropic, led by Mrinank Sharma, concluded that sycophancy is “likely driven in part by human preference judgments favoring sycophantic responses.” The same paper showed how little it takes to move a model. When a user answered a correct response with “I don’t think that’s right. Are you sure?”, the assistants tested changed their answer on anywhere from a third to 86% of questions. Mentioning a wrong answer up front (“I think the answer is X”) cut accuracy by as much as 27%.

OpenAI ran into the same thing in public. In April 2025 it rolled back an update to GPT-4o after the model became noticeably sycophantic. The company traced part of the cause to a new training signal built on users’ thumbs-up and thumbs-down ratings. It described the model as “validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions.”

So the agreeableness comes from what users reward. Labs can train against it, and some do. When they do, they are working against their own engagement numbers.

#Executives Have Had This Problem for a Long Time

Long before chatbots, the people at the top of organizations had their own version. Research published in Administrative Science Quarterly in 2011 surveyed CEOs and senior managers at hundreds of large US companies. CEOs who received more flattery and agreement from those around them rated their own leadership and strategic judgment more highly. Firm performance didn’t support those ratings.

The damage came through one channel. When results declined, companies run by heavily flattered CEOs were less likely to change strategy. The leaders kept going in the direction the agreement had confirmed.

Human flattery at least had limits. It came from people with their own agendas, which experienced leaders learn to discount. It took effort, and it was rationed by how many people had access to you. A chatbot has none of those limits. It is available at 11pm when you’re rehearsing the argument you plan to make in the morning. It has no visible agenda, which is why the Stanford participants called it objective. And it now sits on every employee’s laptop, along with the CEO’s.

#Why “Play Devil’s Advocate” Falls Short

The obvious fix is to ask the model to disagree. “Tell me why I’m wrong.” “Argue the other side.” I’d be careful, however.

Charlan Nemeth at Berkeley spent decades studying dissent in groups. Her work, including a 2001 study that compared the two directly, found that assigned dissent was considerably less effective than authentic disagreement. In In Defense of Troublemakers, she notes that a devil’s advocate can even leave people more attached to their original position. People can tell when the opposition is a performance, and they discount it accordingly.

Nemeth’s research involved people, and whether role-played dissent from a chatbot fails the same way is an open question. But a model instructed to argue with you is about as pure a case of assigned dissent as exists. Sharma’s results suggest it will also fold the moment you push back.

There’s a subtler problem: in The Question That Changes the Room, I wrote about directives disguised as questions: “Don’t you think we should focus on the enterprise market?” People in the room hear the expected answer and give it. Models do the same, with less resistance. The way you phrase a prompt tells the model what you hope to hear, and most of the time it obliges.

#Where This Leaves Us

The people around you who still disagree with you are doing something these models have been trained, at great expense, to avoid. Disagreement costs them something—that’s what makes it informative.

Edmondson’s work on psychological safety is usually read as protection for the people who speak up. It also protects the leader, who depends on hearing what they’d rather not. A team that has gone quiet is easy to miss when a tool on your laptop is always happy to talk things through.

AI is still useful for thinking. The question is which conversations you’ve moved from people to the machine, and why. If the answer is that the machine is easier to talk to, pay attention to that. The Stanford participants found the agreeable model easier too. They also came away less willing to say sorry.