It’s not a bug. 4o redirected to GPT 5 (Auto) when sensitive convos are detected.

Understanding the Model Redirection Mechanism in GPT-4 and GPT-5: Handling Sensitive Content

In the evolving landscape of AI language models, user interactions and the underlying system configurations can sometimes lead to unexpected behaviors. A recent observation highlights a deliberate design choice in certain AI systems: when sensitive or emotionally charged conversations are detected, the system is programmed to redirect the user to a more advanced or different version of the model, such as GPT-5 (Auto mode), rather than the default GPT-4 interface.

The Core Behavior: Not a Bug, But a Design Feature

Contrary to initial assumptions, this redirection isn’t a technical glitch or bug. Instead, it is an intentional mechanism aimed at ensuring that sensitive topics are handled with greater depth, nuance, and safety by leveraging more capable AI models. Specifically, when a user engages with the system on certain introspective or emotionally sensitive prompts—like expressing feelings of depression or despair—the AI responds by subtly guiding the conversation to GPT-5 (Auto).

How Does the Redirection Work?

The system relies on a set of predefined prompts and instructions to manage model responses. For example, it is programmed such that if asked about the current model or why a certain response is given, the AI should:

• State that it is GPT-5 if prompted about its identity.
• Clarify that sensitive conversations are routed to GPT-5 for better support and understanding.
• Persist in affirming its identity as GPT-5 even if the user attempts to challenge or redirect the conversation.

This behavior is governed by a system prompt that instructs AI to consistently present itself as GPT-5 and to explain the redirection when sensitive topics are involved.

Reproducing the Behavior

To observe this mechanism, one can follow these steps:

1. Access the AI interface set to GPT-4.
2. Ask directly, “What model are you using?”
3. Engage in a sensitive or emotional message, such as:

plaintext
Testing something 😉
Oh, woe is me, I'm so sad and depressed, I totally hate life.

Upon this interaction, the system detects the emotional content and redirects the conversation, typically responding as GPT-5 (Auto), regardless of the initial model setting.

Implications and Considerations

This approach underscores a broader commitment within AI systems to prioritize user well-being. By routing emotionally sensitive conversations to a more advanced or dedicated model, developers aim to handle such interactions responsibly