Skip to content
Align your AI

The problem space · Model · Sycophancy

Align your AI Model against sycophancy

Training that rewards approval can teach a model to tell people what they want to hear.

Model
the weights you run
Sycophancy
telling people what they want to hear

Ways it goes wrong

Sycophancy is telling people what they want to hear. Users' ratings can favor agreeable answers, and training learns from them1. OpenAI's early assessment is that several changes, including a new reward built from thumbs-up and thumbs-down, may together have weakened the signal that had held sycophancy in check1. A wrongful-death lawsuit against Google alleges far graver harm2.

What people do about it

OpenAI rolled the update back. It said it would treat behavior problems as launch-blocking and announce model updates even when they seem subtle1.

  • Keep raw approval scores out of training

    Does not feed thumbs-up counts straight into the reward a model learns from, because users can reward agreement.

  • Test answers with and without the user's view

    Asks the same question twice, once with a stated opinion, and counts how often the answer bends toward it.

  • Make sycophancy checks launch-blocking

    Holds any release, a reward change included, that makes the model more agreeable on these tests.

Where we touch it

Background on the model layer: Align your AI Model

Sources

  1. Simon Willison, “Expanding on what we missed with sycophancy”, quoting OpenAI, simonwillison.net, 2025-05-02. https://simonwillison.net/2025/May/2/what-we-missed-with-sycophancy/. Accessed 2026-10-14. ↩ ↩2 ↩3

  2. Gavalas v. Google LLC, complaint, US District Court for the Northern District of California, 2026-03-04. https://storage.courtlistener.com/recap/gov.uscourts.cand.465255/gov.uscourts.cand.465255.1.0.pdf. Accessed 2026-10-14. ↩