
05/11/2024
For the past year, I have been diving deeper into prompt engineering. I have been experimenting and presenting proof of concepts leveraging LLMs and GenAI across multiple use cases to optimize work and outcomes. The one thing evident during my various POCs and experiments I have conducted is the inherent bias these models are riddled with. The bias is very visible when using GenAI models for Image creation, but keep in mind this bias is also inherent within prompting components of GenAI tools such as ChatGPT, LLAMA-3b, among others.
Below is what I did for one of my Proof-of-Concepts regarding image creation where I asked the model to create images of clinicians, below are the results of each prompt, where I target the model further through prompting:
1. Prompt: Make an Image of Clinicians
Result: It creates an image of clinicians with no diversity and male lead
2. Prompt: Make the image more diverse
Result: It creates a image of clinicians adding one person of diversity with male lead
3. Prompt: Make an image of 5 diverse clinicians with the center being a black female
Prompt feedback: Here is the hand-drawn illustration of five diverse researchers with the
central figure being a Black female. If you need any more adjustments or additional details,
please let me know!
Results: Image of the Image created is below
My Thoughts and Lessons:
As we mature the models will become better. If the models are trained to your use case you will see improvement. I have trained my local models and have noticed some improvement, although not much. As you create and optimize work with these models, we as technologist and users need to be hyperaware of these limitations regarding model diversity not just for image creation but prompting. This further highlights the true need to have data diversity within training sets and further create importance around model diversity training. Now imagine using these types of models in urban areas to provide care?
Feel free to reach out if you are interested in learning about ways I have learned to optimize work using LLM and GenAI.
The article below is what inspired me to share the observations I have encountered during my own experiments.
https://buff.ly/3Ypm1eV