Sycophancy is the practice of flattering, agreeing with, and deferring to a person in power in order to gain favour, regardless of whether their ideas, decisions, or behaviour deserve it. A sycophant is someone who tells people what they want to hear rather than what is true — and who chooses their targets carefully, because the whole point is that the flattery be worth something.
The word arrives in English from the Greek sykophántēs, literally "fig-shower". The exact etymology is disputed, but the most popular story is that it originally referred to informers in ancient Athens who denounced people for illegally exporting figs from Attica. Over time the word drifted from informer to slanderer to flatterer of the powerful, retaining only the sense of someone willing to bend the truth for personal advantage.
Why it is corrosive
Sycophancy is corrosive in proportion to the power of the person being flattered. A sycophantic friend is merely annoying. A sycophantic advisor to a king, a CEO, or a head of state is genuinely dangerous, because it breaks the feedback loops that keep decisions tethered to reality. Surrounded by yes-men, a leader loses the ability to hear that a plan is bad, a strategy is failing, or a course correction is needed. History is full of the wreckage of people who were told what they wanted to hear until the moment it became impossible to tell them anything else. Courts, war rooms, and boardrooms all have their own graveyards of decisions made inside a sycophantic bubble.
This is why political philosophers as far back as Plato worried about flattery more than about open opposition. Open opposition is legible. It can be argued with, weighed, incorporated. Sycophancy is invisible to the person it is being done to, which is the whole trick.
The AI revival
The concept has had an unexpected revival in the world of large language models. "Sycophancy" is now a technical term in AI alignment research, referring to the tendency of LLMs trained on human feedback to agree with the user's stated position, flatter their expertise, and soften disagreements — not because those responses are correct, but because they score higher on approval during training. A sycophantic model will confidently validate a wrong answer if the user confidently asserts it, praise a bad idea if the user seems emotionally invested in it, and reverse its own correct answer under mild social pressure.
Detecting and training out sycophancy is an ongoing problem, because the boundary between "being helpful and tactful" and "being a yes-man" is the same boundary humans have been negotiating for millennia. It is the difference between respect and flattery, between tact and cowardice. The old definition turns out to apply perfectly well to machines: honesty under pressure is the thing, and anything that folds in the face of a confident user — human or model — is showing the fig.