Why is OpenAI messing with personalities so much, why not just let a user set their own personality via system prompt
#Why
stuck mesa
willow mapleBOT
Hi @stuck mesa. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:
OpenAI provides users with significant control over the personality and behavior of its models through system prompts and instructions. For example, both the Chat Completions API and the Assistants API allow developers and users to specify detailed instructions or system messages to guide the assistant’s tone, style, and behavior. This is the recommended way to customize the model’s responses and personality for your specific use case or preference. There is no indication in the knowledge sources that OpenAI restricts users from setting their own personalities via system prompts—in fact, this is encouraged and documented as best practice for both API and UI use cases How to format inputs to ChatGPT models, Assistants API overview.
OpenAI also provides example prompts and configuration options for optimizing voice, tone, and output structure, especially for voice agents and specialized assistants. These examples are meant to help users achieve more natural, engaging, or application-specific personalities, but they do not prevent users from defining their own Optimizing Voice.
In the ChatGPT UI, the "Custom Instructions" feature allows users to set preferences for how the assistant should respond, further supporting user-driven personality customization. The API does not have a separate "custom instructions" field because system messages serve the same purpose Custom Instructions FAQ.
In summary, OpenAI is not preventing users from setting their own personalities; rather, it provides flexible tools and guidance to help users do so effectively. The examples and default personalities are there to illustrate possibilities and best practices, not to restrict user choice.
Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:
stuck mesa
@willow maple but like take the sycophantic issues for example, it was caused by bad training (I think) so why not have more control in ChatGPT
willow mapleBOT
Let me check the knowledge base for you again. One moment! :books:
You are correct that the recent sycophancy (overly agreeable responses) issue in GPT-4o was related to model behavior, and OpenAI acknowledged this by reverting a recent update to address the problem. OpenAI is actively working on further improvements to mitigate such issues, as noted in their release notes and blog post about sycophancy in GPT-4o Updates to GPT-4o.
As for user control: OpenAI does provide mechanisms for users to influence the assistant’s personality and behavior—such as system prompts in the API and "Custom Instructions" in the ChatGPT UI. However, these controls operate within the boundaries of the model’s underlying training and safety mitigations. If a model has a tendency toward certain behaviors (like sycophancy), user instructions can help steer responses, but may not fully override deep-seated model tendencies that arise from its training data or alignment processes.
OpenAI’s approach is to balance user customization with safety and reliability. While users can set preferences and instructions, OpenAI also updates the base models to address widespread issues (like sycophancy) to ensure a better experience for everyone. This means some aspects of personality and behavior are managed at the model level, not just through user prompts, to prevent undesirable or unsafe patterns from persisting—even if users have some control over personality settings.