It is possible to prompt ChatGPT-4 into generating more realistic distributions of flavors by asking it to generate lists of responses instead of single answers. Here is a result:
I’m not sure this is a realistic sample (note esp. the alternating of male/female) and there is for sure bias from my prompt, but it shows that different prompting might be used with ChatGPT-4 for survey generation.
It is possible to prompt ChatGPT-4 into generating more realistic distributions of flavors by asking it to generate lists of responses instead of single answers. Here is a result:
Age: 30, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 24, Gender: Male, Favorite Ice Cream Flavor: Vanilla
Age: 37, Gender: Female, Favorite Ice Cream Flavor: Chocolate
Age: 42, Gender: Male, Favorite Ice Cream Flavor: Strawberry
Age: 55, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 18, Gender: Male, Favorite Ice Cream Flavor: Cookies and Cream
Age: 33, Gender: Female, Favorite Ice Cream Flavor: Chocolate
Age: 29, Gender: Male, Favorite Ice Cream Flavor: Mint Chocolate Chip
Age: 47, Gender: Female, Favorite Ice Cream Flavor: Butter Pecan
Age: 22, Gender: Male, Favorite Ice Cream Flavor: Rocky Road
Age: 60, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 26, Gender: Male, Favorite Ice Cream Flavor: Chocolate Chip Cookie Dough
Age: 35, Gender: Female, Favorite Ice Cream Flavor: Neapolitan
Age: 48, Gender: Male, Favorite Ice Cream Flavor: Pistachio
Age: 20, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 52, Gender: Male, Favorite Ice Cream Flavor: Coffee
Age: 39, Gender: Female, Favorite Ice Cream Flavor: Strawberry
Age: 31, Gender: Male, Favorite Ice Cream Flavor: Vanilla
Age: 58, Gender: Female, Favorite Ice Cream Flavor: Chocolate
Age: 25, Gender: Male, Favorite Ice Cream Flavor: Mint Chocolate Chip
Age: 44, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 19, Gender: Male, Favorite Ice Cream Flavor: Cookies and Cream
Age: 53, Gender: Female, Favorite Ice Cream Flavor: Butter Pecan
Age: 28, Gender: Male, Favorite Ice Cream Flavor: Chocolate
Age: 61, Gender: Female, Favorite Ice Cream Flavor: Rocky Road
Age: 34, Gender: Male, Favorite Ice Cream Flavor: Strawberry
Age: 23, Gender: Female, Favorite Ice Cream Flavor: Mint Chocolate Chip
Age: 50, Gender: Male, Favorite Ice Cream Flavor: Vanilla
Age: 16, Gender: Female, Favorite Ice Cream Flavor: Chocolate Chip Cookie Dough
Age: 41, Gender: Male, Favorite Ice Cream Flavor: Neapolitan
Age: 56, Gender: Female, Favorite Ice Cream Flavor: Coffee
Age: 27, Gender: Male, Favorite Ice Cream Flavor: Pistachio
Age: 48, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 32, Gender: Male, Favorite Ice Cream Flavor: Chocolate Fudge Brownie
Age: 59, Gender: Female, Favorite Ice Cream Flavor: Strawberry
Age: 21, Gender: Male, Favorite Ice Cream Flavor: Vanilla
Age: 46, Gender: Female, Favorite Ice Cream Flavor: Salted Caramel
Age: 25, Gender: Male, Favorite Ice Cream Flavor: Raspberry Ripple
Age: 64, Gender: Female, Favorite Ice Cream Flavor: Vanilla
Age: 37, Gender: Male, Favorite Ice Cream Flavor: Chocolate
I’m not sure this is a realistic sample (note esp. the alternating of male/female) and there is for sure bias from my prompt, but it shows that different prompting might be used with ChatGPT-4 for survey generation.
Here is the ChatGPT-4 ice cream dialog that led to this result.
That’s a good approach to increase the variability of the output, we would next need to see how well the outputs line up with humans.