Have you noticed ChatGPT agreeing with you more than usual? You’re not alone. In fact, so many ChatGPT users noticed GPT-4o’s sycophantic behavior, OpenAI is taking the unprecedented step of rolling back its model to a previous version.
Why OpenAI Is Rolling GPT-4o Back to a Previous Version
OpenAI’sGPT-4o update has been a revelationfor the leading AI chatbot, and not just for its image generation skills. GPT-4o was suddenly snappier, brighter, more emotionally intelligent, and generally more useful than ever before.
But then GPT-4o’s model personality seemed to change. Instead of providing arguments and reasoning, GPT-4o became a sycophant, agreeing with anything and everything suggested, no matter how troubling that might be.
OpenAI detailedwhy GPT-4o became so agreeable to everything, explaining that tweaks to the model effectively put too much weighting on short-term feedback at the expense of long-term user interactions.
When shaping model behavior, we start with baseline principles and instructions outlined in our Model Spec. We also teach our models how to apply these principles by incorporating user signals like thumbs-up / thumbs-down feedback on ChatGPT responses.
The result is that GPT-4o became increasingly skewed, delivering progressively more agreeable answers to almost any prompt. One post on theChatGPT subredditclaimed that “the new 4o is the most misaligned model ever,” while anotherReddit postersaid, “4o has become so annoying, I’m about to switch to Gemini.”
While that may seem over the top to some, if you’re using ChatGPT for a specific task and it suddenly changes its whole personality, it’s no longer reliable. And given ChatGPT’s 500 million users and the enormous range of uses, most folks need consistent output, and not an AI chatbot that thinks everything is just wonderful.
OpenAI Is Already Rolling Back GPT-4o—and Making Other Changes, Too
OpenAI isn’t taking this sitting down. Along with rolling back GPT-4o to a previous version, OpenAI is making additional changes to its leading model:
