AI, Imagery, and Unexpected Biases

In my previous post, I shared a frustrating encounter I had at the Starbucks drive-thru. The highlight? A barista who seemed personally offended that I wanted to use Apple Pay and then all but demanded a tip. It rubbed me the wrong way — and like I said, I usually tip! But attitude earns zero.

After writing it up, I thought a related image would really bring the post to life. Before AI image generation, finding a fitting photo was like a scavenger hunt across the internet. And even if I found something close, there was always the copyright cloud hanging overhead. Enter AI art tools — suddenly I could ask for exactly what I needed and have it generated in seconds. It’s not always perfect (hello, baristas with four arms or nightmare-fingers), but when it works, it really works.

So I asked my AI tool to create a cartoon-style image of a barista at a drive-thru window, shoving a tip screen with an attitude. It delivered. But… something was off.

The character it gave me was a Black woman with a scowl.

Now, let me be clear — the race of the barista in my story wasn’t part of the issue. In fact, she was a white Eastern European woman. So when I saw the image, I couldn’t help but wonder: Was that just a random output… or something deeper?

Was this a glitch, or was I looking at something more unsettling — an AI model that had absorbed and regurgitated stereotypes like the “angry Black woman”? Had it unintentionally reflected a societal bias it had learned from the vast sea of human content it was trained on?

That realization bothered me. It wasn’t just about accuracy — it was about fairness. So I asked for the image to be redone with a white woman, thinking it would be a simple switch. But then the new images came back… off. Multiple arms. Distorted faces. Somehow, when asked to portray a white woman with the same level of exaggerated attitude, the system seemed to short-circuit.

Eventually, after a few rounds, I got something close to what I imagined:

Now, I’m not saying the AI is consciously biased (it doesn’t have a consciousness). But it is trained on human-created content — and that content carries all our messy, biased baggage. What we put in is what we get out.

So, was it a coincidence? Maybe. But maybe not.

As AI becomes more woven into our creative and social tools, we have to keep an eye on the assumptions it inherits. It’s powerful — and yes, incredibly helpful — but it’s not neutral. It learns from us, and sometimes it learns the worst of us.

Something to think about the next time we press “generate.”