IntegrAIty: Evaluating Bias in AI

• 

  Home

• 

Inspiration

As AI proliferates, AI systems must be developed with careful consideration of the diverse contexts in which they operate. My goal is to work on making AI more human-centric by identifying and addressing bias in the AI models.

What it does

My web application gives the ability for users to test Gemini’s results for bias in their application domain. For example, a sociologist may look for data like race and ethnic distribution in the generated data and study patterns of inequality, discrimination, and social impacts of biases. Therefore, it is important to reduce the bias as much as possible in the data presented to sociologists. My web application would allow the sociologists to understand the bias in the generated data and correct it before they use it (or use transfer learning or other techniques like semi-supervised learning to tune the model).

How I built it

I used JavaScript, CSS, HTML, and Django to build this website. I started by looking at the API docs for Gemini to figure out how I would implement it in my project.

Challenges I ran into

This is my first time building a website by myself. I found this opportunity on devpost last week, and was inspired to build a project. Learning how to build a website, implement an API, balance my school course load, and deliver results in a week was challenging, but super rewarding.

While making the website, I had a couple issues with Gemini responses to my prompt. I found that unless I was very specific, Gemini had difficulty outputting in a certain format and additionally had extremely strict content restrictors. I also had the problem with prompting GPT, although that one was much more difficult to get responses in a certain format. In addition, because I was prompting Gemini and GPT with multiple queries, I ran out of credits when trying to run on very large batches.

Accomplishments that we're proud of

I am proud that I met the deadline and can have data displayed to users on their own profile page. I am also happy that I am able to compare Gemini and GPT.

This project will also be useful towards my undergraduate research, where I am evaluating bias in LLMs.

What we learned

I learned that the models need more improvement in addressing biases while generating data to be more representative of the real world data.

What's next for IntegrAIty: Evaluating Bias in AI

With more credits, I would love to be able to run more tests at a faster rate so that the data is fairly representative of what a user could expect when utilizing the API. In the future, I will also be working on implementing transfer learning on the Gemini models, so that people or businesses can aim to address the biases for their use case and finetune each API to better fit their needs. I also would like to be able to add semi-supervised learning to reduce bias as shown in results from academic research.

Built With

• css
• django
• html
• javascript