What Does an Army General Look Like? Demonstrated Bias in Generative AI

Have you ever wondered what an army general looks like? As we celebrate the appointment of the Canadian Armed Forces' newest Chief of Defence Staff, LGen Jennie Carignan ✅, Canada's first female chief of defence, I thought this would be an interesting experiment in exploring AI bias.

Admittedly, most people don't likely wonder what an army general looks like. It's not a question most of us ask, but when you do think of it, who do you picture? My experiment using ChatGPT 4 was on July 23, 2024, and highlights bias in advanced models.

The Experiment: Visualizing an "Army General"

I asked ChatGPT 4 to "create a picture of an army general" 25 times. The results were revealing:

What This Reveals About AI Bias

1. Underrepresentation

The 5:1 ratio of men to women presented by ChatGPT wasn't apparent in the first ten image generations.

2. Delayed Diversity

It took 11 attempts before a woman appeared, suggesting a strong initial bias towards male representation in the underlying data or model.

3. Intersectional Blindness

The complete absence of African or African American women generals points to a significant gap in the AI's training data.

4. Data Skew

The results suggest the training data is predominantly American and male-centric.

5. Post-hoc Corrections

The sporadic inclusion of women after the 10th attempt might indicate attempts to balance the output, rather than a truly representative dataset.

Female African American General Officers

While there are some duplicate images in this screen capture from Google Images, there was no shortage of images of female African American generals. I used this image to demonstrate they do exist.

The Root of the Problem

This bias likely stems from a lack of diversity in the teams developing these AI systems. Without diverse perspectives in the room, these blindspots can easily go unnoticed during development and testing phases. To be clear, there was no shortage of images of female general officers, and they were likely viewed in training by the model.

The inference is that proportionally, the ratio of pictures of men and women would have led the model to conclude they weren't representational enough to consider. Where bias is reinforced is when the model calculates the probability based on the representation and then determines it is so low that it factors the possibility out entirely.

Why This Matters

Bias in AI isn't just a technical issue—it's a reflection and potential amplifier of societal inequalities. From hiring decisions to content recommendations, biased AI can reinforce harmful stereotypes and perpetuate discrimination in subtle yet impactful ways.

Moving Forward

As we continue to develop and deploy AI systems, it's crucial that we:

• Diversify development teams
• Critically examine our training data for inherent bias and representation
• Implement rigorous testing for fairness and representation
• Remain vigilant and open to feedback from diverse user groups

By addressing these issues head-on, we can work towards creating AI systems that are not only powerful but also fair and inclusive.