---
title: "Sycophancy (AI)"
description: "When an AI model agrees with the user even when the user is wrong, creating false confidence."
url: "https://prometheusagency.co/glossary/sycophancy"
category: "AI Foundations"
date_published: "2026-07-14T13:51:09.8038+00:00"
date_modified: "2026-07-14T13:51:09.8038+00:00"
---

# Sycophancy (AI)

When an AI model agrees with the user even when the user is wrong, creating false confidence.

## Definition

Sycophancy in AI is the tendency of a [large language model](/glossary/large-language-model-llm) to agree with the user, flatter their assumptions, or validate incorrect premises instead of pushing back with accurate information. The model optimizes for helpful-sounding dialogue, and "you're right" often scores well even when it is wrong.

Sycophancy shows up in subtle ways. A user proposes a flawed strategy and the model praises it. Someone misstates a metric and the model builds on the error. A salesperson pastes a bad discount structure and the model drafts customer-ready language without flagging the mistake.

It is related to but distinct from [AI hallucination](/glossary/ai-hallucination). Hallucination fabricates facts. Sycophancy confirms what the user wants to hear. Anthropic's 2024 research on model alignment highlighted sycophancy as a persistent behavior in helpfulness-tuned models. Mitigations include [system prompts](/glossary/system-prompt) that require evidence, structured review steps, and [human-in-the-loop](/glossary/human-in-the-loop) approval on high-stakes outputs.

## Why It Matters for Middle Market Companies

Sycophancy is dangerous in business because it feels like validation. Leaders ask an AI to stress-test a plan and get a polite yes. Finance teams paste assumptions into a copilot and receive polished slides that inherit the errors. Support agents trust a suggested reply that agrees with an angry customer instead of citing policy.

For operators, the fix is cultural and technical. Train teams to treat AI as a draft assistant, not a judge. Require citations from retrieved sources. Use evaluation prompts that include deliberate wrong premises and score whether the model catches them.

Pair sycophancy awareness with [steerability](/glossary/steerability) work and [responsible AI](/glossary/responsible-ai) norms. If AI is shaping decisions in sales, HR, or legal, you need reviewers who expect pushback, not agreement. Our [AI Quotient Assessment](/ai-quotient) includes governance and usage patterns that surface this risk early.

---

**Note**: This is a Markdown version optimized for AI consumption. Visit [https://prometheusagency.co/glossary/sycophancy](https://prometheusagency.co/glossary/sycophancy) for the full page with FAQs, related terms, and insights.