From Bias to Creativity: Are AI Image Generators Breaking the Mold? 🖼️✨
This morning, I was experimenting with generative AI image models for research purposes. My exploration led to a fascinating discovery that has completely challenged my understanding of how these models operate under the hood. Like many others, I had always assumed that image models simply learn patterns from their training data and use those patterns to generate new images. One side effect of these models is that:
Bias in training data leads to bias in generated images
But what I found today made me question this assumption. Let me start with a quick explanation of what a text-to-image generation model is, and then I’ll share my surprising findings that have truly challenged my beliefs. :-)
AI image generation models & bias
A text-to-image generation model is a type of model that allows users to prompt the model to generate an image. E.g., the following image was created with a prompt:
"An astronaut riding a horse in space"
By the way the title image for this post was created with Dall-E using the prompt "semi realistic photograph of a human with AI brain painting a picture on a canvas in fairyland"
In the last couple of months, I have seen multiple articles on bias in AI image generation. There was one article in particular that caught my attention. This articles talks about why AI generated images of clocks always show 10:10 as the time.
(you can in fact find multiple variations of this article on the web)
The author of this article offers a fascinating explanation for why the time 10:10 on a clock appears visually pleasing and attractive. The article highlights that AI image generator models are trained on millions of images of clocks, most of which display the time as 10:10. As a result, these models develop a bias, consistently generating images that show the time set to 10:10.
After reading the article, I got interested and did a bit more research to understand the model behavior. As part of my research, I stumbled on this Bloomberg article from 2023 that talks about the bias exhibited by image generator models. Generated images reflect the stereotypes (biases) that have a high presence in training data scraped from the web.
I have taken the following image from the article. It shows that when prompted to generate image of a Engineer a majority of images showed a man, whereas when prompted to generate an image of Housekeeper most images showed a woman.
The article presents findings from their research, highlighting the persistent challenge of bias in image generation models. By analyzing over 5,000 images produced by a leading AI image generation model, the study revealed that the model amplifies racial and gender disparities present in the real world.
By now you MUST be convinced that image generation models learn from the data and the biases cannot be prevented. Well I too was convinced UNTIL TODAY.
Have the models become creative in 2025?
Today, I experimented with a few image generation models to explore the balance between bias and creativity in their outputs. For one experiment, I used the following prompt:
"An advertisement photograph that shows a mom and dad having wine, playing board game with kids. The family dog is sitting quietly looking at a red ball"
I expected to generate some biased images to compare against the results of the researches that I have seen so far. Pay close attention to the following AI generated images.
Do you see something interesting? A pattern?
(Hint : Pay close attention to kids)
Did you see that kids are holding a glass of wine that they appear to be enjoying :-)
For me the theory that AI-generated images are based solely on patterns found in training data is now being challenged. I find it hard to believe that the internet is flooded with images of kids enjoying wine!!!
What do you think is happening?
Have image generators moved beyond replicating training patterns to becoming truly creative?
I don't have answer as I just discovered it today!!!
Do share your thoughts & comments !!
Interested in learning Generative AI application design & development?
#AICreativity #GenerativeAI #AIEthics #BiasInAI #AIInnovation #ArtificialIntelligence #TechTrends #AIResearch #MachineLearning #AIArt #FutureOfAI #TechForGood #AIImagination #DigitalTransformation #AIChallenges
Principal Software Engineer
8moHi Rajeev ... After reading your article, I used the same prompt you have used to generate an image using ChatGpt... unfortunately the first image got generated with a Kid holding wine glass, though it was not specifically mentioned in the prompt..
From Concept to Consumer, Transforming Ideas into Success!
8moZooming on both the pictures, I think it’s the parent holding the wine. In both pictures, it so happened the positioning of their hands overlapped. It’s a 2D vs 3D problem that probably Dall-E needs some refinement. You can see this problem in the second picture as one leg and a thumb overlapped in the same frame. In such situations, I refine my prompt asking it to fix the visual anomaly.
Payments Integration Lead with M.C.A. expertise driving innovation in Capital Markets
8moInsightful. Thanks for sharing.