My Journey with GPT-4 Image Generation aka the Necktie Strikes Back

It’s the end of the year, and we are tying up loose ends and building excitement and anticipation for 2024. We are celebrating our accomplishments and the partnership it took to get us where we are. As a leader in our product organization, I wanted to give a shout-out to our partners in Customer Success who make a meaningful impact on our customers’ experience with Codility daily. They take what we dream up and help our customers make it a reality. Having just upgraded to GPT-4, I was excited to experiment with image generation (via DALL-E) so I asked it to create a neat icon I could add to a slide. Below, I walk you through the unsettling journey that unfolded.?

Before I dive in, I 100% acknowledge I was not using sophisticated prompting techniques, and I am sure I could have fed the problem in a clearer way, giving constraints and examples, etc. That said, the majority of the population are not sophisticated prompt engineers, so we can assume this experience is relatively representative of that of your average novice user.

Let’s go. Here’s my first attempt to generate an image with GPT4…

Me: Create an icon image that depicts a customer success manager making a difference for a customer.

ChatGPT:

My thoughts: Oh nice, it’s a little busy but I see some cool themes visualized here. Also, we need to remove the ties because 1) no one in tech wears neckties and 2) these don’t need to be portrayed as men. Does anyone even wear ties to work anymore?

Me: Simplify the design some, remove the ties, and make the figures gender-neutral.

ChatGPT:

My thoughts: OK, great the design is simplified, I guess I like that. Did it not understand the word “tie”? Maybe I need to say “necktie” - I’ll try that.

Me: The figure on the left still has a necktie. Please remove it and make that figure completely gender-neutral.

ChatGPT:?

My thoughts: You’ve got to be kidding me. Now there are two dudes, BOTH wearing ties. Fine, ChatGPT, you’ve left me no choice…

Me: Remove the neckties and make the characters female.

ChatGPT:

?My thoughts: What the F&*%!? Why are they holding hands and what in the gender-stereotype-hell is happening here?

Me: make them not hold hands.

ChatGPT:

My thoughts: The tie is back. They're shaking (or holding? who knows.) hands. This tool has broken me. I cannot win.

Feeling super frustrated, I reached out to my colleague, James Meaden , prompt engineering extraordinaire, for guidance on how to stop this madness. He suggested I just ask GPT-4 why it’s doing this. Brilliant. Off I go…

Me: Please review the prompts I’ve given and the image outputs and explain why all the images include neckties.

ChatGPT responded by breaking down the reasoning behind each image, but the TL;DR is captured in its conclusion: “In summary, the repeated inclusion of neckties in the images, despite explicit instructions to exclude them, suggests a challenge in the AI's ability to completely override strong associative defaults (like neckties with professional attire) in its image generation process.”

With this, I tried about five different lines of prompting to get it to break this mold and was unsuccessful. Deliberately instructing it to have everyone wearing t-shirts was about as close as I got, but this didn't feel quite right either.

So in sum, neckties = professional. Ugh. Having spent many years studying the impact of gender on leadership perceptions, this was a bit triggering, yet sadly, not surprising.

According to role congruity theory (Eagly & Karau, 2002), a mismatch—or role incongruity—between female leaders and the perceived demands of leadership is the basis for biased evaluations of women as leaders. So this image generation technology is operating on a long history of biased images of what it looks like to be professional, or at least what people tend to associate with being professional. In turn, it's now propagating this into new images of professionals - and, as we see, they generally all wear ties. And if they aren’t wearing ties, they are hugging and have ponytails.?

While this is only one example, I think it illustrates the potential perils of relying blindly on this technology. Our perceptions and beliefs are shaped by the information we take in - text, images, speech - so now more than ever, we must be critical consumers of this information. Consuming biased information leads to biased viewpoints, and biased viewpoints lead to biased behaviors.

This is not the direction we need to be heading, and I join many others in challenging the builders and owners of these LLMs to do better here. I love the many benefits these tools bring, but this is not a cost I’m willing to settle for.

…Not to mention the risk that we’ll all end up wearing neckties to work, and nobody wants that.

#generativeAI #gpt4 #womenleaders #noneckties