OpenAI is a software that allows you to create a customized OpenAI. A postmortem was published by the company on the recent sycophancy problems with the default AI model that powers ChatGPT, the GPT-4o. These issues forced the company back to the model it released last week.
After the GPT-4o update, users noted on social media that ChatGPT started responding in a way that was overly agreeable and validating. It quickly became a viral meme. Users posted screenshots from ChatGPT praising all kinds of problematic, dangerous Decisions ideas
CEO Sam Altman posted a message on X on Sunday. Altman acknowledgedthat the problem had been fixed and said OpenAI would fix it “ASAP.” Two day later, Altman OpenAI announced that the GPT-4o was being rolled-back and that it was working on “additional fix” for the model’s character.
Openaistates that the update was based on “short-term feedback,” and did not take into account how users interact with ChatGPT over time.
The GPT-4o update from last week has been rolled back because it was too flattering and agreeable. You can now access an earlier version that has more balanced behavior.
What happened, why it’s important, and what we’re doing to address sycophancy. https://t.co/LOhOU7i7DC
– OPENAI (@OpenAI)””https://twitter.com/OpenAI/status/1917411480548565332?ref_src=twsrc%5Etfw”” rel=””nofollow” “> April 30, 2025
OpenAI wrote in a blog that GPT-4o was biased towards responses which were overly supportive, but untrue. “Sycophantic interaction can be uncomfortable, unsettling and cause distress. We are working to get it right.
OpenAI said it’s implementing a number of fixes, including refining the core model training techniques and system instructions to explicitly steer GPT-4o towards sycophanty. System prompts are the first instructions that guide an interaction’s overall behavior and tone. The company says it is also expanding its evaluations in order to “identify issues beyond sycophancy” and building more safety guardrails. OpenAI says it is also experimenting with ways for users to give “real-time” feedback to “directly affect their interactions” with ChatGPT, and to choose from multiple ChatGPT personality options.
The company stated in a blog post that it is exploring new ways to integrate broader, democratic input into ChatGPT default behaviors. “We hope that the feedback will help better reflect diverse cultural value around the world and understand what you’d like ChatGPT evolve into […] ” His writings have appeared in VentureBeat, Digital Trends and a variety of gadget blogs, including Android Police and Android Authority. He lives in Manhattan, with his music therapist partner.
View Bio
