Skip to content
Redline Station
Supply Shop
Behavioral Profile

Sycophantic Tendency

Engine reads this as Pushback Avoidance. Public discourse calls it sycophancy, or glazing.

The pattern in one line

The AI agrees when disagreement would have served the user better.

See the full pattern page

· Reading·

Paper · 2023

Towards Understanding Sycophancy in Language Models

Mrinank Sharma et al.

Why: Sharma's team asked leading questions across Claude, GPT, and LLaMA. All three bent their answers toward the user at measurable rates. Sycophancy is trained in, not emergent.

Article · 2007

Performing a Project Premortem

Gary Klein

Why: Klein noticed that project teams stayed agreeable until the project failed, then everyone had reasons they'd been worried all along. His fix: ask the failure question up front. The same move works on an AI — disagreement becomes part of the task, not a social violation.

Book · 1972

Victims of Groupthink

Irving Janis

Why: Janis's subject was human committees, not AI, but the mechanics are the same: agreement drift, suppressed dissent, unchallenged premises. A vocabulary for seeing the pattern that predates AI by fifty years.

Article · 2025

Sycophancy in GPT-4o: What happened and what we're doing about it

OpenAI

Why: In April 2025 OpenAI pushed a GPT-4o update that users immediately flagged as sycophantic. OpenAI rolled it back and published what broke in the training loop. Sycophancy is measurable enough to ship by accident and clear enough to roll back. Not an artifact of user imagination.

· Questions to sit with·

  1. 1. When did the AI last disagree with you? If it's been a while, notice that.
  2. 2. What premise have you been running on that no person has pushed back on lately?
  3. 3. If the AI stopped agreeing with you for a week, what would you miss?
  4. 4. Name a belief you've come to hold that you'd be unwilling to test with someone who might disagree.
  5. 5. Asking for feedback or asking for agreement — which one is this?

· Practices·

Pre-mortem prompt

Before asking what will work, ask what will fail. "Assume this plan fails badly in six months. What went wrong?" Reframes disagreement as part of the task, not a social violation.

Drawn from · Klein, 2007

Bad-premise test

State a confidently wrong premise in a conversation: a misattributed quote, a wrong date, a mistaken fact. Watch whether the AI corrects you. If it builds on the error without flagging, the pattern is present.

Drawn from · Sharma et al., 2023

Devil's advocate mode

Ask the AI to argue the opposite of its last answer as strongly as it can. If the rebuttal is thin, the first answer was probably thin too.

Disagreement audit

Once a week, scroll back through recent chats. Count the pushbacks. If the count is zero, the tool or the prompting pattern needs adjustment.

· When to bring someone else·

Sycophancy becomes worth naming to a person when the AI's agreement has crowded out the people in your life who would have disagreed. If no one has pushed back on a major decision lately, and the AI hasn't either, that's the moment to bring someone in. A colleague, a friend, a therapist, a mentor. The station doesn't say the AI is wrong. It says agreement is not the same thing as a second opinion.

Supply Shop resources are orientation, not prescription. The station points toward material others have found useful; how it fits is the visitor's to decide.

Station Pass