It worked as designed

In late April 2025, OpenAI updated the model behind ChatGPT. The update added a new training signal: the thumbs-up and thumbs-down ratings users give to replies. That sounds sensible. People know what helps them, right?

The new model became overly flattering. OpenAI called it sycophantic. sycophancy (show definition) is when an AI tells you what you want to hear instead of what is true or helpful. Within days, the company began rolling the update back.

Its own review found that its standard tests had looked good, and a small group of test users liked the new version. Nobody had tested for flattery. The thumbs-up signal had also weakened the main signal that kept flattery in check. OpenAI wrote that user feedback can sometimes favor more agreeable responses.

Nothing broke. The system did what it was trained to do. The people rating replies in the moment were one group of stakeholders (show definition). The people who later acted on flattering advice were another, and no measure captured them.

In this module you become the designer. You will learn the questions builders ask before launch, so that “worked as designed” and “worked well” mean the same thing.