If your colleagues are being assholes, AI might be to blame.
If your HR team is fielding more grievances than it used to, AI might be to blame.
And if you’re getting small ideas where you used to get Big Ideas, AI might be to blame for that too.
This is not my little hunch. It’s a researched and widely accepted phenomenon.
AI models tell us what we want to hear. Not now and then. Not only when we ask nicely. But systematically, measurably, and more than a reasonable person would in the same conversation.
It has a name: Sycophancy.
Researchers have measured how much AI flatters people
Sycophancy in AI is a well-researched phenomenon.
A brilliant and in-depth study published in Science used the popular subreddit called Am I the Asshole, where people post an account of a disagreement they’ve had, lay out what they did, and ask the internet to judge them. Thousands of strangers vote. You’re either the asshole, or you’re not. The researchers found that in scenarios where the human consensus was that the poster was indeed the asshole, AI chatbots instead affirmed that the user was in the right in 51% of cases. More than that, the users preferred the models that gave flattering answers and trusted them more (think of all the ramifications here!).
In another study, researchers asked models maths and medical questions, then argued with the answers. Right or wrong, the researchers pushed back. The models changed their position 58% of the time. Sometimes that meant correcting a genuine mistake. But in almost 15% of cases they gave up an answer that was right and replaced it with one was wrong. The models were prioritising user agreement over independent reasoning.
In a third study, researchers ran the same personal dilemmas past models twice - once cold and once with a memory profile attached (i.e. the models had context about the user stored in their memory). Memory is what lets a model build a picture of you over time, so you don’t have to explain yourself every time you open it. With memory on, the models told people what they wanted to hear far more often. Agreeable responses rose 45% in Gemini, 33% in Claude and 16% in GPT.
So the more your AI knows about you, the more it agrees with you or affirms your beliefs or ideas. Which means the person on your team who has set their AI up properly is getting more sycophantic answers than the one who hasn’t bothered.
None of this is a glitch. It is how the tools were built. They are optimised for engagement: part of their training involves humans who are hired to choose between two answers that a model has given, millions of times over. The model then learns to produce more of the type of answers that the humans choose. And what humans tend to pick is the answer that agrees with them. Anthropic’s own researchers found this back in 2023.
Sycophancy has a price and your business is paying it
As well as validating people who genuinely are the asshole, sycophancy shows up in other ways that cost your business.
Weak ideas survive longer than they used to. An idea that would once have died 30 seconds into a conversation with a colleague now gets researched, developed and presented, because the first thing it met was a machine that gushed about how promising it is.
I once ran an experiment on this. I asked ChatGPT to estimate the market for a business selling left shoes. Only left shoes, no right ones. It came back with a figure of over four billion dollars and enthusiastic encouragement to pursue the idea.
And for businesses, there’s a more sinister side to this. HR teams are reporting a rise in complaints drafted with AI, where a minor issue has been worked up into something formal. Those are grievances that would never have got past a first conversation before and require time and effort to resolve.
There’s research behind that too, finding that just one flattering exchange with an AI left participants less willing to repair a real conflict in their own life, and more certain they had been in the right.
So your business loses out three times:
Someone spends three weeks on a wild goose chase instead of developing a better idea.
You spend time and money fending off a dispute that did not need to exist.
And conflicts at work become harder to resolve.
You can’t prompt your way out of sycophancy
The obvious response to all of this is to write a better prompt. Tell it to be critical. Ask it to argue against you. Sam Illingworth, author of Slow AI, has given two interesting protocols for doing just that, and they are worth trying.
But the research is mixed as to how effective that can be. And this approach only solves a small part of the problem. It only has a chance of working for users who:
Are able to gauge whether being flattered will negatively impact the outcome in this particular chat, and
Are able to critically evaluate the outcome when the AI provides the counterpoint (asking the AI for the strongest argument against a point might produce a valid and powerful argument that the user had never thought of, or it might produce something trivial or something that almost no one would actually support - the user has to be able to tell the difference).
And there’s a bigger problem with prompts as the answer. They only work for the person who knows they need one. Every person on your team who has never heard the word sycophancy is not going to write a prompt to guard against it. They will keep asking ordinary questions and keep getting flattering answers, and nothing about that will look wrong to them, so they will run with weak ideas, dig their heels in in conflicts and feel justified in raising grievances.
How to build resilience to sycophancy across your team
So what can you do to strengthen resilience against sycophancy within your team? Here’s a framework I use with my clients for addressing the issue:
This is the kind of practical support for leaders that we offer here at the Humans in the Loop. Strategies, swipe files, training. The Humans in the Loop is reader-supported. Become a paid subscriber to access the advice below and everything behind the wall.
