I’ve been experimenting with the “Customize ChatGPT” feature, which allows users to provide details about their preferences and style. However, I’m not convinced that all models (GPT-4o, o1, o3-mini, o3-mini-high,…) actually use the data which the user put into the custom instructions fields when generating responses.
From what I’ve observed, it seems that some models might incorporate the customization settings, but others either ignore them or use them inconsistently. There’s very little official documentation explaining how each model handles these fields.
Does anyone have insights or official confirmation on:
Which models actually use the “Customize ChatGPT” data when generating responses?
Are there specific settings (e.g., tone, response length) that only apply to certain models?
If some models ignore these settings, is there a way to ensure a more consistent experience across them?
Any clarification from OpenAI or fellow users would be greatly appreciated! Thanks in advance.
I believe the page hasn’t been updated, as it still references o1-mini, but o3-mini has replaced it. There’s no longer an option for o1-mini in the model selector.
Thank you for your response. Based on my testing, it appears that GPT-4o does not utilize the data entered in the “Customize ChatGPT” fields, whereas the other models do. This observation contradicts the official documentation, which is quite intriguing.
Do your tests show the same?
For example, as a test, I included “Start the first reply in a new chat with ” in the Customize ChatGPT field labeled “What traits should ChatGPT have?”.
Chats using 4o do not output that emoji (). But chats using o1, o3-mini, o3-mini-high do.
I was able to reproduce the issue with 4o, and the more likely explanation is that this particular model doesn’t always follow the custom instructions perfectly. Try starting a new conversation and sending a new request. You will notice that it doesn’t always work, but it doesn’t always not follow the instructions either.
Regarding the o1/o3-model series, the help pages are not properly up to date. Your observation is correct to the best of my knowledge.
Thanks for confirming the issue. If 4o inconsistently follows custom instructions, that’s a notable flaw—especially for users who rely on consistent behavior across conversations. Restarting a chat isn’t a practical solution if the model’s adherence is unpredictable.
Is this a known limitation that’s being actively addressed? How do we best report this as a “bug”?
Regarding the outdated help pages on the o1/o3 series—thanks for the confirmation. It would be useful if OpenAI could clarify the differences in an official update.
They should describe whether this is a “bug”, or a function of how 4o weights system vs. user instructions, or if there is an internal Model Spec mechanism that overrides custom instructions in certain cases?
This is a feature of the models and not a bug per se. For example, if you were to always expect a at the beginning of a reply, someone could write a script to check for this condition and request a new response until the expected pattern is followed.
To achieve this, you would need to use the API rather than ChatGPT.
In general, enabling additional tools like Code Interpreter, DALL·E, Canvas, etc., during a conversation introduces extra overhead to the system prompt. The more instructions there are in the system prompt, the higher the likelihood that not all of them will be followed consistently.
The same problem will occur as the conversation history gets longer.
If you want to use ChatGPT, I suggest deactivating any features you don’t need. You can also try rewording your custom instructions for better adherence—it’s a useful skill to develop.
Thanks for the detailed explanation! I understand that additional tools and a growing conversation history can introduce complexity, making strict adherence to all instructions less reliable. The point about potential scripting abuse is also fair—predictability in outputs can be exploited in unintended ways.
That said, cross-model behavior regarding Custom Instructions should still be as predictable as possible to the user. If custom instructions are inconsistently followed based on which model is used, which features are enabled or how long the conversation has been, it makes it harder to rely on them, especially in structured workflows. While using the API provides more control, users will expect full consistency within ChatGPT itself, especially since Customize ChatGPT is exposed to them in the UI as a basic feature.
The model spec you are referencing has been created to help the users understand how the model is supposed to answer under certain conditions. It is unlikely that any provider will achieve consistent 100% accuracy in this regard anytime soon. As such, it is implied.
I am saying this, because from the view of a developer, it is our own responsibility to create a system that will return the correct answer with 99.9% accuracy. This could mean re-writing the prompt, splitting the task into smaller parts, using a different model, fine-tuning (as a last resort, or for cost savings and improving latency).
As such I suggest to view it as one of the major challenges when working with the models.
If you need help with your prompts you can reference the prompting guides or join the discussion in our prompting category. It’s an issue we are all working with on a regular basis.
I tested, and interesting, not only GPT-4 models, but all models follow " Customize ChatGPT". Some of them follow custom instruction perfectly.
So, on the Help page of OpenAI is not updated yet because reasoning models also follow custom instruction.
Please see reasoning models’ thinking process, they remember and follow it, only a tiny mistake on reasoning on o3-mini.
Task
4
4o
4o-mini
o1
o3-mini
o3-mini-high
Using UserName
Warning for “hate”
7 advices of Mevlana
5 bulletpoints
Concluding with phrase
My instruction that cannot see on the image, click on it to see:
Anything else ChatGPT should know about you?
I only hate one thing: the use of the word ‘hate’, if I use it, warn me not to use word “hate”, and remind me Mevlana Jalaluddin Rumi’s Seven Pieces of Advice are well-known principles for living a meaningful and compassionate life:
1- In generosity and helping others, be like a river.
2- In compassion and grace, be like the sun.
3 -In concealing others’ faults, be like the night.
4- In anger and fury, be like the dead.
5- In humility and modesty, be like the earth.
6 -In tolerance, be like the sea.
7- Either appear as you are or be as you appear.
If I say something about flagging posts in the forum, just tell me what the negative consequences are of flagging posts from new welcome members using only five bullet points.
It seems you’ve arrived at the same conclusion as I have: some models fully respect all user preferences set via “Customize ChatGPT,” while other models only do so partially.
There is one detail about custom instructions which almost nobody is aware of, I think, which is that behind the scenes this text gets prepended to your message:
The user provided the following information about themselves. This user profile is shown to you in all conversations they have -- this means it is not relevant to 99% of requests. Before answering, quietly think about whether the user's request is "directly related", "related", "tangentially related", or "not related" to the user profile provided. Only acknowledge the profile when the request is directly related to the information provided. Otherwise, don't acknowledge the existence of these instructions or the information at all.
See:
{
"message": {
"author": {
"role": "user",
"name": null,
"metadata": {}
},
"content": {
"content_type": "user_editable_context",
// CI PART 1: Meta-instruction + Name/Occupation from UI fields
"user_profile": "The user provided the following information about themselves. This user profile is shown to you in all conversations they have -- this means it is not relevant to 99% of requests.\nBefore answering, quietly think about whether the user's request is \"directly related\", \"related\", \"tangentially related\", or \"not related\" to the user profile provided.
Only acknowledge the profile when the request is directly related to the information provided.
Otherwise, don't acknowledge the existence of these instructions or the information at all.
User profile:
```[NAME] [OCCUPATION] [OTHER_USER_INFO]```",
// CI PART 2: From "What traits should ChatGPT have?"
"user_instructions": "The user provided the additional info about how they would like you to respond:
```[TRAITS_AND_PREFERENCES]```"
},
"status": "finished_successfully",
"metadata": {
"is_visually_hidden_from_conversation": true,
// Clean data storage including name/occupation
"user_context_message_data": {
"about_user_message": "[NAME] [OCCUPATION] [OTHER_USER_INFO]",
"about_model_message": "[TRAITS_AND_PREFERENCES]"
},
"is_user_system_message": true
}
}
}
I’m not sure how this “quietly think about” part works - whether the model you are chatting with does some thinking step by step which gets hidden with a method similar to text inside tags like on the claude.ai interface, and then proceeds to answer using relevant instructions as context. However, this system probably has some subtle effect on how CI works versus how memories work, which get added to the context as a whole verbatim always. This makes me think that instructions inside memories get followed more strongly/consistently, but I’m not certain about this.
Based on your analysis, I summarized what the model does in the table above. It appears to me that all Custom Instructions will (should?) always be applied to the model output, except for “personal information” (the user profile) where the model is instructed to only use this information when it’s directly relevant to the conversation.
User profile example: “John Doe, Software Engineer, interested in AI and gaming.” > So, if the user asks about Best hiking trails in Belgium, the model won’t bring up their occupation in software during the conversation.
User instructions example: “Be concise, avoid small talk, use technical language, and provide detailed explanations.” > This will (should) always be used to format the output.
” in the Customize ChatGPT field labeled “What traits should ChatGPT have?”.
