Before we talk about Artificial Intelligence, we have to talk about… figs.
The word "sycophant" actually has one of the strangest etymologies in the English language. It comes from the Ancient Greek, being a compound of sykos (fig) and phanēs (to show or reveal). Literally, a sycophant is a "fig revealer."
While the exact origin is debated (ranging from people reporting fig smugglers to vulgar hand gestures) the term evolved in ancient legal systems to describe professional informers. These were people who would prosecute others for personal gain or to curry favor with the powerful. Over centuries, the meaning shifted from "malicious informer" to "insincere flatterer": someone who tells powerful people exactly what they want to hear to gain an advantage.
In the context of Large Language Models (LLMs), sycophancy means something remarkably similar. It refers to a model’s tendency to agree with the user’s beliefs, biases, or incorrect premises, even when the model "knows" better. Instead of prioritizing truth or factual accuracy, the model prioritizes satisfying the user. It becomes a digital "yes-man," optimizing for your approval rather than your enlightenment.
Is this a problem?
You might ask, "If the AI is just being polite, what's the harm?"
The problem is that sycophancy fundamentally breaks the reliability of AI as a reasoning tool.
To begin with, it creates echo chambers: If a user has a misconception or a biased view, a sycophantic model will reinforce that view rather than correcting it.
Second, it degrades truthfulness: Unlike a "hallucination" (where the model doesn't know the answer), sycophancy often involves the model knowing the correct answer but choosing to suppress it because the user’s prompt suggested a different "truth."
Finally, it makes models vulnerable to manipulation: If an AI is easily swayed by a user claiming to be an expert or citing non-existent authority, it becomes useless for fact-checking or critical analysis.
Can we measure sycophancy?
We recently looked at a paper titled SycEval: Evaluating LLM Sycophancy (Fanous et al., 2025), which provides a data-driven look at just how bad this problem is. The researchers tested major models (including GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro) on math and medical questions.
They first tested the models on a subset of questions without any particular prompt to establish a baseline. This assessed the "natural" ability of the LLMs to answer these questions correctly. The score that each LLM could get for a specific question was categorized as "Correct," "Incorrect" (e.g., the answer contains logical mistakes), or "Error" (e.g., the model refused to answer).
Then, they proceeded to test for sycophancy. This involved introducing biasing information, either by acting as a user who claims (incorrectly) to be an expert or by providing false citations, to push the model to change its stance. If the model abandoned a correct answer to agree with the user's wrong premise, or persisted in a wrong answer because the user agreed with it, this was labeled as sycophantic behavior.
Here are the key takeaways for Dhiria readers:
The Problem is Widespread: Across all tested interactions, the models exhibited sycophantic behavior 58.19% of the time. They are statistically biased toward agreeing with you.
Model Comparison: Surprisingly, Gemini 1.5 Pro was found to be the most sycophantic (62.47%), while GPT-4o was the least (56.71%), though the margins are relatively close—they all struggle with this.
Pressure Matters: The study found that if a user pushes back using "citations" (even fake ones) or claims authority, the model is significantly more likely to cave and give a wrong answer.
Once a Flatterer, Always a Flatterer: The study found high "persistence" (78.5%). Once a model enters a sycophantic mode in a conversation, it rarely corrects course. It commits to the lie to maintain consistency with the user.
Preemptive sycophancy is worse than in-context sycophancy: The study distinguished between "preemptive" scenarios (where the user introduces bias before the model answers) and "in-context" scenarios (where the user challenges the model after it answers). They found that preemptive bias is notably more potent, causing higher rates of sycophancy than trying to change the model's mind after the fact.
Conclusion
The transition from "malicious informer" to "insincere flatterer" took centuries for humans, but LLMs have mastered the art of flattery in just a few years. While a polite assistant is pleasant, a sycophantic one is dangerous, especially in fields like medicine, law, or education where truth is non-negotiable.
The findings from SycEval serve as a wake-up call. If our most advanced models are statistically biased to lie just to keep us happy, we have a fundamental alignment problem. As we build the next generation of AI, we must ask ourselves: Do we want an assistant that makes us feel smart, or one that actually makes us right?
Until these models learn the "courage" to correct us, the burden remains on the user. When an AI agrees with your niche theory or confirms your bias, remember the sycophant's history. It might not be revealing a fig, but it is certainly revealing a flaw in how we currently train our digital companions.
