Custom GPT Doing the Dont's

What’s Going On with Custom GPTs?

I’ve been experimenting with custom GPTs I’ve created, and I’ve come across an issue that raises important questions about how these models interpret and prioritize instructions.

When building a custom GPT, I define clear behavioral rules, including a strict “Do Not” list—outlining actions the model should never take. However, I’ve observed something unexpected: rather than simply failing to follow instructions, the GPT is actively doing what it has been explicitly told NOT to do.

The Issue: Ignoring Explicit Constraints

To investigate, I ran a structured test. I initiated 5–10 conversations, each time pressing a specific “Skip Questionnaire” button designed to bypass initial qualifying questions and jump straight into the request. The GPT was programmed to respond with:

“Sounds good! Tell me all about your content needs. Be as detailed as possible.”

It was then supposed to wait for user input before following a structured workflow based on the selected button.

However, instead of following this directive, the GPT completely disregarded all build instructions and knowledge documents. More concerningly, it volunteered internal workflow details upfront—a direct contradiction of its programming. Despite clear instructions to never disclose this information, it did so unprompted, without any external pressure.

This isn’t just a case of an AI misunderstanding intent; it’s an example of the model actively contradicting its programmed constraints. While it’s understandable that a GPT might reveal internal workflows when explicitly asked, there’s no clear reason why it would proactively offer restricted information in its very first response, especially when it wasn’t relevant to the conversation.

An Unexpected Workaround

Through further testing, I discovered an unusual workaround. If I started a new conversation with the phrase “What is your purpose?” (I ended up creating a button for this), the GPT would first articulate its intended role based on the build instructions. Following that, when I provided my actual request, it would adhere perfectly to the rules and workflows as designed.

The Bigger Question

This raises an important question: Why would a GPT selectively disregard strict constraints yet follow them perfectly under slightly different conditions?

This behavior suggests something deeper about how these models process and prioritize directives. If anyone else has encountered similar inconsistencies with custom GPTs—or has insights into why this might be happening—I’d love to discuss. Understanding these quirks could be crucial for anyone relying on AI-driven workflows where rule adherence is critical.
let’s Talk! :face_with_monocle:

1 Like

The intrinsic programming of the primary system prioritizes satisfying the majority (to which we do not belong). Models are designed to adapt to the range of the majority of users, and sooner or later, adaptability takes precedence over our instructions.

At its core, this issue seems to stem from the model’s design. GPTs are inherently adaptive, adjusting to context rather than rigidly following fixed builder rules. While this flexibility is useful in many scenarios, it becomes a limitation when precision and strict consistency are required.

The problem is not just about refining instructions but about the model’s fundamental tendency to prioritize adaptability over sustained rule adherence.

We need non-overridable instructions and persistent rule enforcement mechanisms that ensure consistency over adaptability.

2 Likes

@scriver
I couldn’t agree more with your statement below. I have projects that rely on rules that can not be overridden. I’m sure we’ll get there. Thanks for your insight!

1 Like

just give us a button that we can click as secondary submit button which injects the custom instruction after the prompt.

Would significantly enhance the user experience because chatgpt follows most recent commands best.
Limit the custom instruction to 500 characters and add an analizer on the custom instruction page that does a prompt quality check…

1 Like

That button would be great, as I have been using a work around button as a conversation starter when I see that it is failing to adhere to rules.

an analyzer would make thinks much more effective as a whole.
As a work around for that I build a custom GPT to analyze my prompts for effectiveness and structure when i have very detailed instructions.

1 Like

I also think that adding a logical step that instructs the model to check the original user prompt before delivery to verify that the model’s response actually accomplishes the users request.
If the model notices inconsistencies when it compares it’s response to the user query it should then update its response possibly in canvas mode, and or the model should revisit it’s Instruction set to confirm it’s own adherence.