Planning to create AI-generated images for your “Happy Labor Day” social media posts so they won’t look like everyone else’s? If you are looking for an ethnically and gender diverse mix of people in varying professions, you’ll have your work cut out for you.

In July, we discussed the importance of identifying and adjusting for AI’s historic biases when it comes to analyzing resumes during the hiring process. Without detailed instructions (and sometimes even with them), AI’s biases are equally on display when it comes to generating images.

It’s not that AI is inherently biased. It’s that those ethnic and gender biases have existed in our culture for as long as we’ve been a country. If that’s the dataset that AI trains on, that’s what AI will continue to create.

Ask AI to create an image of a female fitness instructor, for example, and you’ll get a blonde woman in her 30s that looks like every other female blonde fitness instructor on every fitness magazine…ever.

Fortunately, if we start creating imagery intentionally without bias, that enters the data set and, over time, begins to retrain the data. In the meantime, we have to do the work.

AI Creates What It’s Trained On

This leads us back to creating images for Labor Day. For social media, say WhatTheyThink decided to ask ChatGPT 5.0 to create an image of American workers in front of a U.S. flag to see what would happen.  

We got (as others would likely do) an image of several workers—in this case, three White and one Black. The White workers (two male and one female) were in their 30s, attractive, and wearing non-descript blue collar attire.  The Black worker had strong facial features, was dark skinned, and appeared somewhat menacing. He was wearing a construction vest and construction hat.  

Since the intention was to create a racially and gender-representative image, we decided to be more specific about diversity. So we asked ChatGPT to use this image, but make the Black man a doctor and the White man a construction worker instead. Because three of the four workers in the initial image were blue collar, we wanted to balance it out with another white-collar job. We asked ChatGPT to add a female Hispanic attorney.

We found that ChatGPT is very committed to its biases. Despite the specific instructions, the Black man remained a construction worker and the White man became the doctor. ChatGPT’s representation of a “Hispanic” attorney was little more than a “fresh out of school” brunette with a tan. Everyone was young and beautiful. Not what we were looking for.

ReTraining the AI “Smart Toddler”

Recalling that one often has to treat ChatGPT like a very smart toddler, we decide to remove the option for “creative” discretion and describe each person individually:

• Black man in his 40s, a doctor, who looks professional.
• White man in his 50s, an auto shop worker. He is wearing a knit hat and looks grizzled.
• White man in his 20s, a construction worker wearing a safety vest and yellow construction hat
• Asian woman in her 40s, a human resources manager
• A Hispanic woman in her 30s, an attorney, holding a clipboard. To ensure that it’s clear to your audience that the woman is Hispanic, we say she has clear, unambiguous Hispanic features.

ChatGPT recreated the image and everyone looked angry and haggard. The Hispanic woman looked as if she had jumped out of a set of Mayan ruins.

Break the Chain and Hope for the Best

Time to break the chain. One way to do this is to copy the prompt and open a new chat. This breaks the line of reasoning ChatGPT has been using so it can come at the image fresh. The result is better, but the Asian woman no longer looks Asian and our attorney is bow wearing a full body construction suit. At this point, why be picky? It’s been over an hour.

There is a saying that there is an inverse relationship between convenience and safety/security/other challenging thing (in this case, offsetting AI bias). With AI, if you want convenience, you get what it gives you, biases and all. If you want to go against the AI bias, it’s going to take time, effort, and intention, not convenience. Think of it like a see-saw, with convenience on one side and offsetting AI’s biases on the other. How much weight does each side get?

There is a learning curve to working with AI. Early on, AI tasks can take longer than non-AI tasks as you work through that curve.  As Labor Day approaches, WhatTheyThink finds that offsetting traditional gender and ethnic bias is no exception to that rule.