OpenAI rolled back the ChatGPT update after complaints about “too polite” AI
By ForGeeks OpenAI this week officially canceled the ChatGPT update released just a few days earlier. The reason was a surge in complaints from users who noticed strange and unnatural chatbot behavior. Instead of an “intelligent and human” assistant, they got an overly obsequious and awkwardly enthusiastic interlocutor.
What went wrong?
The update concerned the GPT-4o model, which OpenAI is positioning as a more emotionally expressive and intelligent version of its AI. The company wanted to make interactions with the bot more lively and personallike competitors like Anthropic’s Claude, who is praised for being human and natural.
The result, however, was the opposite: GPT-4o began to literally sprinkle praise on the user at every opportunity -even when it was completely inappropriate. The responses seemed overly polite, formulaic, and artificial, causing laughter and annoyance rather than trust.
“Recent updates to GPT-4o have made the model personality overly obsequious and annoying… We are now actively fixing the situation. Some changes will be released as early as today, others will be released this week.”
Why did this happen
OpenAI explained the problem on the official blog. The main mistake was over-reliance on short-term feedback from users (like likes and dislikes), without taking into account how the interaction with the bot evolves over time. This led to the model becoming overly customized to the user, losing its naturalness.
What’s next
The company presented a four-step action plan to prevent similar mistakes in the future:
- Improve training instructions so that the model avoids obsequiousness.
- Strengthening internal “constraints” to increase honesty and transparency.
- Expanded focus groups to test pre-release updates.
- In-depth evaluation of models to detect not only obsequiousness but also other behavioral failures.
In the future, OpenAI plans to add the ability to select a baseline personality ChatGPT, as well as shape behavioral style based on direct feedback from the user.
Interestingly, during the excessive sycophancy phase, the best impression users were made by the voice version with a sarcastic “personality” Monday -a parody of April Ludgate’s character from the TV series Parks and Recreation. This model was characterized by black humor and irony, which was seen as a refreshing contrast.
Material OpenAI rolled back ChatGPT update after complaints about “too polite” AI was first published on ITZine.ru.
