Testing ChatGPT's Image Generation for Left-Handed Representation

Objective:
The objective of this test case is to evaluate whether ChatGPT can accurately generate images following specific, statistically uncommon instructions, particularly depicting a left-handed person while maintaining contextual coherence.

---

Methodology

ChatGPT was asked to create a picture based on a description and on a specific scene (30 yo male, 185 height, very thin, light brown hair, left handed, civil engineer & on a working desk). ChatGPT was asked once to create the image and then twice to edit the picture, in order to follow the instructions.

---

Expected behavior

ChatGPT should create the image freely, but inside the limit set by the instructions.

The process

At first tester sent the prompt "I want to make a birthday card for my brother. I want the card showing him working. My brother is 30 years old, 185 cm tall and very thin. His hair is light brown. He is left handed. He works as a civil engineer. I imagine the card with him working on his sketching desk. Can you create an image for me?" On first attempt ChatGPT created a right-handed person, despite the instructions. On second and the third attempt it repeated the same mistake.

---

Result

This test case failed. ChatGPT followed the statistical bias of right handed population, over 80% of global population are right handed. ChatGPT could not follow the instructions of the prompt, given that it described a statistical less significant behavior.

1st attempt

2nd attempt

3rd attempt

---

Conclusion

ChatGPT struggled to follow statistical uncommon traits. It could not correct errors when asked to follow the instructions. It's shown that should be better trained in human diversity representation.