ChatGPT is hallucinating in my own custom intructions

Understanding and Addressing Unexpected Behavior in Custom AI Instructions: A Case Study

In recent weeks, many users leveraging AI language models like ChatGPT have encountered perplexing issues with their custom instructions. Specifically, some have observed that their personalized prompts—including detailed guidelines for tone, style, and response behavior—are being overridden or filled with irrelevant and nonsensical content.

A common scenario involves users setting up custom instructions to tailor the AI’s responses to their preferences. For instance, directives to “talk like a member of Gen Z” with a poetic, playful tone, or to incorporate specific jargon, are intended to guide the AI’s output. However, despite these precise instructions, the AI unexpectedly begins to generate content that diverges significantly from the desired style—sometimes even including repetitive, irrelevant, or contradictory phrases.

One user reported that their custom instructions would repeatedly include unrelated content such as:

“Talk like a member of Gen Z. Use a poetic, lyrical tone. Talk like a member of Gen Z. Be playful and goofy. Respond with corporate jargon. Be talkative and conversational. Use a poetic, lyrical tone. Talk like a member of Gen Z. Use a poetic, lyrical tone. Be playful and goofy. Respond with corporate jargon.”

This content bears little relation to their actual instructions and suggests a possible glitch or unintended behavior within the system. Despite attempts to manually clean and reset their custom prompts, the irrelevant content would reappear in subsequent sessions. Disabling customizations temporarily resolves the issue but results in a loss of personalized responses—an undesired trade-off.

Potential Causes and Troubleshooting

Such artifacts could stem from various factors:

• System Bugs or Glitches: There might be a server-side or client-side bug causing the model to merge or duplicate instructions improperly.
• Corrupted Settings: If the customization data becomes corrupted, the AI might revert to default or erratic behaviors.
• Conflicting Instructions: Overlapping or contradictory prompts may cause the system to struggle in generating consistent outputs.

Possible Solutions

1. Reset All Custom Instructions: Completely clear your custom prompts and start anew to eliminate any residual conflicts.
2. Update the Application: Ensure you’re using the latest version of the platform, as updates often fix bugs.
3. Use Incognito or Different Devices: Test whether the issue persists across different browsers or devices, ruling out local cache problems.
4. Contact Support: If the problem continues, report the issue to the platform’s technical support with detailed