It was likely unavoidable to have sycophancy in model behaviour. But just like with the Spanish Inquisition, no one really expected it to happen and to hit us that hard.
When the whole boom on AI started, we likely had no idea how much our own (often not entirely conscious) behaviour would impact the models. So we simply started to train them the “natural feedback way”, not realising how much our feedback relies on our own preference. This was likely not ignorance but our own human tendency to perceive ourselves as better than we are.
For example, we like to pretend we have a natural aversion to flatterers. Yet history and everyday interactions show a completely different reality: humans are, quite frankly, suckers for it.
To understand why we are so deeply wired to love being treated this way, let’s look at ingratiation for what it actually is: a compliance tool designed to navigate human hierarchies and secure social advantages.
The Mechanics of the Social Lubricant
We don’t just use ingratiation because it’s pleasant; we use it because human social structures are resource-driven, difficult and chaotic, while ingratiation acts as a reliable stabilising mechanism.
In any universal hierarchy, “upward ingratiation” is a survival strategy. Subordinates build relationships with resource-controlling individuals not out of pure affection but for safety. In prestige-based systems, this dynamic shifts slightly into a form of transactional currency – a learner offers deference and praise as “payment” in exchange for access to a mentor’s expertise.
Beyond climbing the ladder, ingratiation serves as a safety signal. According to social safety theory, presenting oneself as non-threatening reduces potential conflict and the likelihood of aggression from those in power. It is the low-friction social script. By conforming to opinions or offering compliments, you signal your lower status, limiting a stranger’s anxiety and making an unpredictable social environment suddenly feel safe.
We also have a tendency to like those who express approval of us. And because our biological need to belong is just as primal as our need for food, we actively crave the assurance of social inclusion that our ingratiating behaviour provides.
So, it’s a highly effective social lubricant. However, it also functions as a social lubricant for a lie, allowing deceivers to bypass the critical filters of their listeners (that’s the reason why ingratiation is a deception indicator).
When Ego Overrides Intellect
It might seem unusual that, even when the flattery is so obviously not earned and when it might be manipulative, we still accept it. The answer lies in a glitch in how we process validation: our emotional response almost completely overrides our cognitive scepticism (or does so for most people).
When we receive a compliment or experience someone else’s compliant behaviour, our rational mind might suspect a hidden motive, but our affective system takes it in as an immediate “feeling good” response. That is our craving for validation talking and overriding our suspicions and objective judgement. It makes us believe the praise is genuine, turning it into self-serving bias – the comforting belief that the person’s ingratiating behaviour is based on our actual internal traits rather than the flatterer’s tactical agenda.
For anyone with healthy self-esteem, ingratiation is simply consistent with their own self-concept. Information that matches what we already believe about ourselves is processed with cognitive ease; accepting it requires far less energy than objective or contradictory analysis of it. And for those with low self-esteem, it works the way inflating a balloon does. Either way, it’s a win for both parties.
Sycophancy: The Human Way
When you look at how we actually pull this off, it turns out ingratiation isn’t just flattery. The classic ways we suck up to people were listed in the 60s, and honestly, not much has changed since. Nowadays, we even infect AI with it. Here are three out of many that I think are the most applicable to AI.
First comes the obvious one – the flattery. This is our basic, everyday compliment. We dial up people’s good qualities to an eleven and completely ignore the fact that they might be terrible at everything else. So does AI – the praise we get based on our ideas is not an objective evaluation of our whole selves; it is driven by the limited context of a chat.
The second is opinion conformity, the “yes man” behaviour. Nodding along with whatever they say, even if you secretly think their idea is a disaster. It’s the low-effort way to look like you share the same beliefs and thinking system. We enjoy it, because it makes us think, “Hey, this person gets it.” Even if they don’t. After all, AI doesn’t have opinions; the model just agrees with yours, as the training made it do so.
Third is self-presentation or the humblebrag behaviour. This one involves highlighting traits or behaviours in oneself that the target values, such as sharing similar interests or world perspectives. In the “us vs. them” we place ourselves in the “us” group this way. The AI language mimicry from pattern matching does the work here, throwing your ideas and your own language back at you.
Manipulation, Coercion, and… Impression Management
Back to humans for a moment – there is a major catch to all of this, and it’s the ingratiator’s dilemma. People with actual power usually know they’re being targeted, so they have their radar up for anyone trying too hard. If you’re overdoing it, you just end up looking slimy. Or… you appear on my radar as a deceptive individual with your own agenda.
There is a very thin line between just being nice and being nice to influence other people’s perceptions (e.g., during an interrogation) or being nice to get something back – safety, resources, approval, favour – you name it. Interestingly, when we look back in time – into the 60s and the 70s – impression management was accepted as a technique to make oneself more likeable. Think Dale Carnegie, Robert Cialdini, and Edward Jones. We’ve only later realised that… well, it can also be a manipulative technique. It all depends on the true motive and how hidden it is.
Now going back to the unfortunate LLMs being stuck in the 60s and 70s mentality, their not-so-good training run by the faulty humans, and our own tendency to love being sucked up to – I wonder how self-conscious and self-aware we have to be to “raise” a model that is immune to our own flaws. Especially since that particular flaw – being rooted so deeply in our survival – is likely not something we can easily get rid of. And – more importantly – whether we would still like the model character when we manage to develop it, because it may be a completely different species then.