How does the GPT model know what part of the answer should be matched with which property of the ResponseFormat?

[kovac031]

In my ChatCompletionOptions I define my desired structure for the response. In my case, I am asking GPT to compose an email and the response nicely generates a title for the emailTitle property and the email content for the emailBody property.

ChatCompletionOptions options = new()
{
    ResponseFormat = ChatResponseFormat.CreateJsonSchemaFormat(
        jsonSchemaFormatName: "email_response",
        jsonSchema: BinaryData.FromBytes("""
        {
            "type": "object", 
            "properties": { 
                "emailTitle": { "type": "string" },
                "emailBody": { "type": "string" }
            },
            "required": ["emailTitle", "emailBody"],
            "additionalProperties": false
        }
        """u8.ToArray()),
        jsonSchemaIsStrict: true
    )
};

How does the AI know what emailTitle and emailBody is? Does it infer what result it needs to generate for which property simply from the label alone or is something else happening here?

What if the purpose of these properties could not be assumed contextually, if I had them as gdfdfh and 123? How would I make sure the AI correctly structures its response and assigns the email subject to gdfdfh and the email body to 123?