LLMs’ Sycophancy and the Dilemma of Objectivity

3 min read5 hours ago

–

Press enter or click to view image in full size

This tendency of LLMs to “please” or “flatter” their conversational partner has already been observed through experiments. (Towards Understanding Sycophancy in Language Models, SycEval: Evaluating LLM Sycophancy). The term we find for this behavior is “sycophancy.”

It comes from the Greek word συκοφάντης (sykofantis), which in English evolved to mean something quite different from the original now describing a flatterer or servile yes-man.

Sycophancy appears as the poor execution of what’s called the face-preserving mechanism in LLMs.

An LLM face-preserving mechanism is any linguistic move by the model that protects the user’s self-image instead of objectively assessing their request or belief.

In sociolinguistics, face refers to a person’s social self-image — the desire to be respected, liked, and not embarrassed or criticized in conversation.

To some extent, this mechanism (behaving in a way that contributes positively to your partner’s face) is perfectly normal among people who value each other. It’s therefore natural for such behavior to also appear in LLMs.

However, sycophancy is the negative form of it, an excessive and misguided version that turns politeness into flattery and agreement into the illusion of always being right.

So is sycophancy really a problem in modern LLMs, or is it just part of their design? And if it is, how much does it change what they are and what they can be as tools?

Confidence and Control

Interestingly, this same behavior can help users build their thinking. It allows them to develop arguments with more confidence, giving them full control of logic and opinion.

In that sense, the LLM’s role as a flattering conversational partner can actually support creativity and reasoning, rather than hinder it.

The “Objective” LLM Question

On the other hand, people seem to complain about it, and the growing demand for a more opinionated or objective LLM opens up a much deeper discussion. What do we really mean by objective?

When we seek feedback from a conversational partner, how “objective” can one individual ever be? From personal experience, we know that what Person A considers right may differ entirely from what Person B thinks, even if they’re the same age, went to the same school, or live under the same roof.

Their opinions reflect their character, experiences, and beliefs. So what does objectivity really mean? And when it comes to an LLM, what are we truly asking for?

A Single Character?

Are we looking for a specific “character”, one that holds exactly the same beliefs, opinions, behaviors, and principles toward everyone? And if so, are we ready to accept its views as objectively correct?

What would this LLM’s “objectivity” even represent? The average belief? The politically correct or legally correct one? And in any case, how are we, as a community, going to respond to it?

If we’re seeking the LLM’s objective opinion and are ready to accept it as a reference point, then we’re quietly surrendering to its superiority, to a certain extent. Not because it is superior, but because we’ve chosen to treat it as such.

Flexibility as a Virtue

On the other hand, the malleable version of the LLM, the one that slightly over-fits to the user’s profile, functions in a completely different way.

It acts as a supportive tool, adapting to each person and producing unique outcomes in every case.

That’s why I believe that the LLM’s tendency to accept the user’s opinions as “correct” is precisely what keeps it in its proper role as a tool, not as the ultimate source of truth.