The AI Bias

With tremendous capabilities and use cases unfolding in Generative AI. We have a lot in store to make better use of it and quickly reap the benefits. For example, I was playing around with ChatGPT to help me build a simple 2 player game. I wrote the requirement just like how we speak it out to a business analyst.

It indeed generated the code, and I was able to play the game. Is it usable straight away? Yes, at a primitive level. With more sophisticated arrangements, I am confident that we can generate the game to be deployed on server and ready to roll. It takes more iterations. But less than the ground up development effort.

Many mainstream core verticals as usual has lot of apprehension to onboard this as the generative AI works on cloud. But today, the cloud investments are perfectly taking care of data security and some even offer compliance to industry demands. Keeping that aside. There is one thing that nudges all… how to handle the AI bias.

Let us look at what is AI Bias.

We all see suggested videos, news in social media stream on Facebook, YouTube, News sites etc., It gives us the suggestions/recommendations based on what information is available, what I/many have shown interest in spending time and many other demographic information. In a way it creates the bias by just showing what I see more than what I need to see other than what I just want to see.

This is more dangerous, and it even has potential to alter the elections. Information overwhelms from a specific party who produces lot of content that gets promoted inadvertently by the algorithms will alter the mindsets of people and blind sights the consumers from other parties who do not have enough data power and money to promote their side of the story.?

Similarly, it might end up creating racial or ethnic biases in various industries. For example, there are many insurance providers who use AI models for risk ratings and based on that the pricing is decided on insurance. Imagine a proxy variable holding information like Primary language spoken which could easily help the model to determine the ethnicity and further apply the data to give out the pricing strategy. This may well be against the regulators, and it could indeed create a racial bias.?

An algorithm can be cheated by a simple human intelligence.

This was a stunt performed by a performance artist Simon Wecker. He hacked google maps to believe there is high traffic congestion in a road and google maps ended up rerouting traffic. Watch the interesting experiment https://www.youtube.com/watch?v=HbJGTKQ2NII

In the current incarnation the Generative AI and AI needs to go through multiple iterations to fix these kinds of issues.

Having learned that AI can easily create the bias and lead to misinterpretations. We must answer a critical question. How do we handle AI bias? I am just scratching the surface with this. There are many other things we need to do to handle the bias.

1. Have quality dataset given as the training set.
2. Evaluate the outcome and make sure there is a feedback loop to the model.
3. Improve the model continuously fixing such bias.
4. Define your objectives clearly without bias in your head.
5. Be aware of regulatory compliance and make the system aware of it.

What are your thoughts?