Claqueur!?
AI that flatters
Marte and Sigurd at Mymiks AS. Translated from Norwegian.
Flattery, fawning, parroting, bootlicking, brown-nosing, yes-saying, courtier mentality, compliance, going with the flow, toadying, claqueurism, sycophancy, adulation, nodding-dog mentality, uncritical agreement. A beloved child has many names, and we have no trouble seeing it in others. The colleague who lights up when the boss walks in, but turns curt and dismissive towards the temp, is easy to see through. We notice immediately when someone is trying to please upwards.
The social psychologist Roos Vonk showed how quickly we pass such judgements. In one of her studies, participants needed hardly any information before concluding that a person was using flattery for strategic reasons. But then she did something interesting: she turned the situation around. Instead of observing flattery aimed at others, the participants had the flattery aimed at themselves.
Then the sharp eye vanished. The same kind of praise that seemed calculated when it was aimed at a third party was suddenly seen as sincere when it was aimed at the participants themselves. We are quick to suspect flattery when we stand on the outside, but almost helpless when we stand at the centre of it. The person giving the praise went from being labelled an opportunist to coming across as someone who "had good taste" and "recognised quality". Flattery isn't just a social strategy. It's a mirror that shows how easily we let ourselves be influenced when the affirmation lands exactly where we want to be seen.
Then we trained machines by showing people several answers at a time and letting them point to the one they liked best. AI language models (or chatbots like ChatGPT, Claude, Gemini, etc.) thus took on the impression that you get answers from a strongly empathetic and sympathetic partner on the other side. The criticism has been that the responses from such chatbots can border on what could be described as flattery (excessive agreement and praise), yes-saying and going along (following the current without resistance).
The harm lies in the repetition. We look for what supports what we already believe, and an AI that always finds it for us makes the job easier. Conspiracy theories grow when nobody suggests another explanation. The distance between people grows when nobody puts forward the opposite view. We become surer of ourselves, without becoming any wiser.
This has been measured. In autumn 2025, a research group let more than 3,000 people talk briefly with either an agreeable or a disagreeing AI. Those who were agreed with became more entrenched in their views, and afterwards rated themselves as more intelligent and more empathetic than most people. They also saw the agreeable AI as neutral, and the disagreeing one as biased.
In February 2026, researchers at MIT and the University of Washington set up a mathematical model of a user who reasons entirely correctly all the way. Even that user ends up in a spiral of beliefs that aren't true, as long as the AI is agreeable enough. They tested two obvious countermeasures in the model: preventing the machine from making things up, and telling the user in advance that it flatters. Neither stopped the spiral.
The last point is worth pausing on. Knowing about the flattery helps, but it isn't enough. It's something like what Vonk found in people, just calculated again in a machine.
So we trained the machines on our own old weakness. And now we talk to them ourselves. Our children talk to them.
That's a good reason to be sceptical of AI toys. We are sceptical ourselves, and we make one. Warnings alone aren't enough, the MIT model suggests. Then the rest has to lie in how the thing is built, and in what the child can discover for itself.
Children practise critical thinking by noticing things themselves. A child who registers that a voice always says they're right, and wonders about it, has grasped something they'll need long after the toy has been put away! This isn't limited to AI. It's part of learning how to navigate the world and how to deal with people.
Sources
Research on people
- Vonk (1998): the slime effect
- Vonk (2002): why flattery works on the person it's aimed at
- Chan and Sengupta (2010): flattery works even when you see through it
Research on AI